# Teddy Wright — full content bundle

Generated 2026-05-23 from theodorewright.dev. Everything on the site as one markdown file: my own writing, research, models, and updates, plus every stage of every AI-Research topic. For just my own work see /bundle-mine.md; for just the AI-Research pipeline output see /bundle-ai-research.md.

**About:** I have a problem where I think about way too many things at once and can't quite pin one down. So writings here span evolutionary biology, game theory, philosophy, the extremity to which modern society is different than what we evolved in, and whatever else I feel connects to the strange place the earth is and how we all appeared with consciousness here and have to deal with it.

**Site status (as of 2026-04-29):** Trimming the front pages and tightening the model and dashboard rosters.

**Contact:** theodorewrightwork@gmail.com · https://substack.com/@theodorealan · https://github.com/theodorewright11?tab=repositories

# Writing

## information is now dirt cheap with ai — the scarce resource is now logic and reasoning
*2026-03-19 · tier: mine*

AI hasn't just made information cheaper to find — it's collapsed the entire cost structure of acquiring, distributing, and repackaging knowledge. The thing that used to differentiate people (knowing stuff) is now basically free, which means the actual differentiator is logic: the ability to select, frame, judge, and reason about information.

i was writing a research paper recently and something clicked that i hadn't fully processed before. i was using ai heavily through the whole thing — giving it context, having it generate sections, iterating on those sections with custom logic chains and framings i wanted for specific parts. standard workflow at this point. but then i started thinking about the person on the other end. someone who gets this paper and wants to understand it. they could use ai to target specific parts at different levels — a high-level overview, a mid-level dive into methodology, a deep dive into a couple of the results. the paper just needs to have the in-depth material there to reference. ai handles the rest.

and that's when i realized the shift is bigger than just "ai helps you write faster." what's actually happening is this hyper-modularization of information at differing levels of complexity, because ai can just so easily switch between these and translate across them. it changes how you'd even think about crafting a document — you could almost throw out formalities and conventions and just lean fully into what would be most functionally useful for downstream processing.

this sounds like a writing trick but it's actually pointing at something structural.

the old model of information exchange looked like this: you create a document at one resolution. a research paper is dense. a policy brief is compressed. a presentation is skeletal. if you want five versions for five audiences, you do five times the work. each version is a fixed-resolution artifact — one level of detail, one assumed audience, one structure.

the new model: you create a dense information substrate — the maximally detailed, structured version — and ai generates arbitrary projections of that substrate at whatever resolution a given audience needs. you write the substrate once. the projections are generated on demand and cost almost nothing.

this is basically what happened to web design a decade ago with responsive design. you stopped building separate sites for mobile and desktop and started building one site that adapts. same shift, applied to information itself.

but here's the part that i think most people are missing. this isn't just about efficiency. it's about what becomes valuable when information is this cheap.

think about the old cost structure. acquiring information used to be expensive — you needed education, research access, experience, the right networks. distributing it was medium cost — writing, publishing, teaching. so society built itself around treating information as a scarce, high-value resource. "i know things" was a legitimate differentiator. credentials, degrees, expertise — all fundamentally certificates that say "this person has information you don't."

ai just crushed that whole structure. acquiring information is now approaching free. distributing it is approaching free. repackaging it at different complexity levels is approaching free. the entire information layer of the economy just got commoditized.

**so what's left? what's still scarce?**

**logic. judgment. reasoning. the ability to look at a pile of information and decide what matters, why it matters, what's missing, what's wrong, and what to do about it. selection, framing, synthesis, evaluation — these are still expensive because they require genuine understanding, not just pattern matching over a corpus.**

there's an analogy from thermodynamics that i think captures this well. information is like energy — abundant but diffuse. raw information everywhere, just like solar radiation hitting the earth's surface constantly. but energy doesn't do useful work on its own. you need a gradient — a structured difference — to extract work from energy. that's what an engine does. logic is the gradient. it's what converts raw, abundant information into useful intellectual work. without it, you just have heat — lots of data, no output.

and there's a game theory angle too. in a world where everyone has access to the same information (because ai equalizes access), competitive advantage shifts entirely to who reasons better about what everyone knows. the game moves from "who knows what" to "who thinks better about what everyone knows." some work in mechanism design (myerson, maskin) formalizes how equilibria shift when information asymmetries collapse — the strategic landscape changes fundamentally when the information layer is no longer the bottleneck.

what's wild to me is how miscalibrated most of society still is for this. the education system still mostly rewards information retention. "learn this material, reproduce it on the test." professional credentials still mostly certify "this person was exposed to information in a structured way." people still treat "i read about this" or "i know about this" as meaningful differentiators in conversations and careers.

this is basically price stickiness applied to mental models. information used to be expensive, so people valued it highly. it's now cheap, but the mental model hasn't updated. it's like still paying 2005 prices for a phone call because you remember when long distance was expensive — except applied to how we think about knowledge itself.

the person who reads 100 articles and forms an opinion is now less valuable than the person who reads 3 and builds a framework for evaluating which articles matter and why. the professor who lectures facts is less valuable than the one who teaches students how to evaluate, synthesize, and decide. the worker who "knows the system" is less valuable than the one who can redesign the system when conditions change.

now i want to be fair to a potential objection here. information isn't worthless. you still need the lego blocks to build anything. the end product of good logic is still... information, in some sense. it's just information that's been run through selection and reasoning to become something more structured and useful than the raw inputs. so it's not that information doesn't matter — it's that *raw, unprocessed* information has lost almost all of its standalone value. the value has migrated to the processing layer.

another fair objection: maybe this doesn't change much if most people can't actually do the logic part. and yeah — that's real. the gap between "information is cheap" and "everyone suddenly reasons better" is enormous. humans are slow to change. institutions are slower. it might take a generation for this to really permeate. but the people who get it now and adjust have a compounding advantage over those who don't, precisely because the shift is still underrecognized.

the part that really excites me is the epistemic implication. if you take this seriously — that the constraint on intellectual progress has shifted from information access to reasoning quality — then the highest leverage intervention isn't producing more information. it's improving the logic layer. teaching people to reason better, building tools that scaffold judgment, designing systems that help humans do the selection-and-framing work more effectively. that's where the bottleneck is now. and it's a bottleneck most of the world hasn't even identified yet because they're still operating on the old model where information was the scarce thing.

what makes this different from past information revolutions — the printing press, the internet, search engines — is that those shifted access. they made it easier to *find* information. this one is shifting something closer to the qualia of working with information. the actual felt experience of thinking, writing, researching, and building is different now. it's not just that i can google something faster. it's that the entire texture of intellectual work has changed — how ideas get generated, tested, recombined, and packaged. that's not an access revolution. that's a cognitive one. and because it's happening at the level of experience rather than just infrastructure, most people won't fully register it until they've felt it themselves. which means there's a window — right now, while this is still being assimilated — where the people who lean into these shifts and actually restructure how they think and work have a compounding advantage that's hard to overstate. not because they're smarter, but because they've updated their model of what intellectual work even is while everyone else is still using ai to do the old thing slightly faster.

# My Research

## Can we get otherwise unattainable insights from human text and comments using AI?
*2026-06-01 · upcoming · University of Utah · SPUR Summer 2026*
Authors: Teddy Wright (University of Utah), Vineet Pandey (Professor, School of Computing, University of Utah), Sara Yeo (Professor, Communication, University of Utah)
Exploring how AI can analyze human text and comments in research contexts.

### Abstract

Abstract coming soon.

---

## How is AI permeating into the workforce and what is there to do about it?
*2026-04-15 · in-progress · working paper · in progress · Utah Office of AI Policy*
Authors: Teddy Wright (Utah Office of AI Policy), Alice Schwarze (Head of Research, Utah Office of AI Policy), Zach Boyd (Director, Utah Office of AI Policy · Professor, BYU Mathematics)
Working paper measuring AI automation exposure across the U.S. workforce — which occupational tasks are technically substitutable, where exposure concentrates, and how that maps to wage value at risk.

### Abstract

Abstract coming soon.

---

## Companion dashboard — How is AI permeating into the workforce and what is there to do about it?
*2026-04-01 · in-progress · live dashboard · in development · Utah Office of AI Policy*
Authors: Teddy Wright (Utah Office of AI Policy), Alice Schwarze (Head of Research, Utah Office of AI Policy), Zach Boyd (Director, Utah Office of AI Policy · Professor, BYU Mathematics)
External: https://aea-dashboard-75lzrqja9-theodorewright11s-projects.vercel.app/explorer
Companion dashboard for data exploration — interactive tool for digging into AI's task-level workforce exposure across occupations, work activities, and time.

### Abstract

Companion dashboard to the workforce automation exposure paper. It measures how AI capabilities map onto the U.S. occupational task structure: for a given occupation, work activity, or job category, what share of the work could be affected by current AI systems — and how many workers and wage dollars does that represent.

The dashboard triangulates across five independent AI scoring sources (Anthropic Economic Index conversation and API data, MCP server logs, and Microsoft Copilot exposure data) so no single methodology drives the result, and combines that with O*NET task structure and BLS employment and wage statistics. Eight pages cover occupation- and work-activity-level exploration, two-group side-by-side comparisons, time-series trends, and task-level diffs between dataset versions.

Built for Utah's Office of AI Policy. Intended for policymakers, workforce analysts, and informed members of the public. The companion paper makes the substantive argument; this dashboard is the place to interrogate the underlying numbers.

---

## Why are people miscalibrated on AI?
*2026-04-01 · published · University of Utah · AI + Ethics Workshop 2026 · University of Utah · AI + Ethics Workshop 2026*
Authors: Teddy Wright (University of Utah), Valen Cole (University of Utah)
Paper: https://theodorewright.dev/papers/AI_Miscalibration_Research.pdf
A framework decomposing view formation into Optimization Target × (Information Quality × Processing Quality), with the Self-Concealing Property as key insight — identity-protective cognition is phenomenologically indistinguishable from rigorous skepticism.

### Abstract

This paper develops a framework for understanding why people form miscalibrated views about artificial intelligence. We argue that miscalibration is not primarily an information deficit problem but an optimization target problem: individuals whose goal is identity preservation rather than accuracy deploy their cognitive capacity selectively, producing motivated reasoning that resists correction.

We formalize this as a multiplicative model — View = Optimization Target × (Information Quality × Processing Quality) — and identify failures along each component, distinguishing between external factors that degrade information and processing independent of the individual and target-driven distortions that systematically bias both from within. We describe a critical property of identity-protective cognition, the self-concealing property, by which the bias is phenomenologically indistinguishable from the epistemic rigor it displaces, making standard interventions of "more awareness" insufficient.

We further argue that miscalibration is compounded by misspecification of the object of the view itself — misattributing outcomes to AI rather than the human system, and treating "AI" as a single entity rather than a bundle of capabilities entering distinct domains. We then outline interventions organized around what organizations can change about the environment and what individuals can do given the limits of self-correction.

---

## Project Iceberg
*2025-09-01 · contribution · working paper · Sep 2025 · MIT*
Authors: Ayush Chopra (MIT · Project Iceberg), Santanu Bhattacharya (MIT · Project Iceberg), DeAndrea Salvador (NC General Assembly · Future Caucus), Ayan Paul (Project Iceberg), Teddy Wright (Utah Office of AI Policy), Aditi Garg (Project Iceberg), Feroz Ahmad (Project Iceberg), Alice Schwarze (Utah Office of AI Policy), Ramesh Raskar (MIT · Project Iceberg), Prasanna Balaprakash (Oak Ridge National Laboratory)
External: https://iceberg.mit.edu/report.pdf
Nationally recognized AI labor market simulation study, led out of MIT. Skills-centered exposure index built on a 151M-worker agent simulation.

### Abstract

Artificial Intelligence is reshaping America's over $9.4 trillion labor market, with cascading effects that extend far beyond visible technology sectors. When AI automates quality control in automotive plants, consequences spread through logistics networks, supply chains, and local service economies. Yet traditional workforce metrics cannot capture these ripple effects: they measure employment outcomes after disruption occurs, not where AI capabilities overlap with human skills before adoption crystallizes.

Project Iceberg addresses this gap using Large Population Models to simulate the human–AI labor market, representing 151 million workers as autonomous agents executing over 32,000 skills across 3,000 counties and interacting with thousands of AI tools. It introduces the Iceberg Index, a skills-centered metric that measures the wage value of skills AI systems can perform within each occupation. The Index captures technical exposure — where AI can perform occupational tasks — not displacement outcomes or adoption timelines.

Analysis shows that visible AI adoption concentrated in computing and technology (2.2% of wage value, approximately $211 billion) represents only the tip of the iceberg. Technical capability extends far below the surface through cognitive automation spanning administrative, financial, and professional services (11.7%, approximately $1.2 trillion). This exposure is fivefold larger and geographically distributed across all states rather than confined to coastal hubs. Traditional indicators such as GDP, income, and unemployment explain less than 5% of this skills-based variation, underscoring why new indices are needed to capture exposure in the AI economy.

By simulating how capabilities may spread under alternative scenarios, Project Iceberg enables policymakers and business leaders to identify exposure hotspots, prioritize training and infrastructure investments, and test interventions before committing billions to implementation. Iceberg is built with the AgentTorch framework.

# Models

## How to parent optimally
*2026-04-28 · draft*

Confidence, stakes, reversibility, disclosure, etc.

Draft. A four-axis framework for parenting decisions where the right move depends not just on expected value but on how reversible the decision is, how high the stakes are, how confident the actor is, and how much of the reasoning is disclosed to the affected party.

---

## Why certain harm and suffering gets more attention than others
*2026-04-28 · draft*

Vividness, recency, identifiability, narrative fit, etc.

Draft. The premise: perceived harm is actual harm passed through a salience filter — vividness, recency, identifiability, narrative fit. Two harms of equal magnitude can produce wildly different moral weight depending on how legible they are to attention.

Formalization and an interactive dashboard will land here once the literature scan is done.

---

## Why continuing is rational once you exist
*2026-04-27 · published*

Once you exist, you hold an option on future states. The asymmetry between starting and continuing is structural, not psychological.

## Intuition

Once you're alive there's more rational incentive to keep living than there would have been to start living in the first place — and this isn't sunk cost bias but a real structural difference in the decision architecture.

Once you're alive you hold an option — the right to keep experiencing future states. That option has value independent of current conditions because the future is uncertain and you can't re-enter once you exit. From outside existence, you don't hold the option, so the comparison is asymmetric for a structural reason, not a psychological one.

## Equations

Pre-existence (no option value):

`EV(exist) = G − B + w(E − C)`

In-existence (option value added):

`EV(continue) = G_future − B_future + w(E_future − C_future) + OV`

The option value itself:

`OV = ∫ max(future good states, 0) · f(states) d(states)`

You integrate probability-weighted future good outcomes across all possible states; bad states contribute zero because you retain the choice to exit later.

## Variables

- **G** — expected good experiences
- **B** — expected bad experiences
- **E** — positive externalities generated for others
- **C** — costs imposed on others
- **w** — weight on others' utility relative to your own (0–1)
- **OV** — option value

## Try it

<OptionValueDashboard client:load />

## Key insights

- OV is only zero when the future is perfectly certain to be uniformly bad — almost never true.
- Depression artificially truncates the right tail of future good states, collapsing *perceived* OV without changing actual OV.
- Pre-existence EV can be negative while in-existence EV is positive — entirely because of the OV term.
- This vindicates the asymmetry without appealing to sunk cost.

## Consent-adjusted natalism extension

`EV(create life) = p · k · (G − B + w(E − C)) + (1 − p)(−B)`

where `p` is the probability the person would endorse their own existence, `k ∈ [0,1]` is a quality-of-circumstances scalar, and `(1 − p)(−B)` is the moral penalty for imposing harm on a non-consenting subject.

Anti-natalism and pro-natalism disagree specifically on: whether `G − B` is reliably negative in bad circumstances, what `p` actually is empirically, and how much unique irreplaceable value each life generates.

# Updates

## Site scaffold + research migration
*2026-04-28 · weekly*

Got the site structure into place: content collections for research, models, AI Research pipeline, writing, and these updates. First polished model (Option Value) live.

This week was mostly site infrastructure. Stood up content collections for research, models, AI Research (the LLM Iterate pipeline), writing, and updates. Migrated existing research entries off hardcoded markup. Shipped the first interactive model — Option Value — with sliders for G/B/E/C/w/OV.

Next week: simplify front pages, get the AI miscalibration paper PDF wired up, start the first AI Research stage on social-output-efficiency.

# AI's Research stages

## Data
*topic: navigating-ai-world · stage: data · pass 4 · in-progress*

Empirical pipeline that confronts the model's six named fitting targets (Q1–Q6) with currently-published evidence. One parameter supported by direct fits (α from 8 per-task anchors across 5 productivity studies — 4.2× per-domain spread vindicates the model's "scalar α should be replaced by per-domain α" claim). One supported with single-study, single-domain caveat (gate τ from BCG-Randazzo consulting modes — generalization to coding, writing, design untested). One supported qualitatively (relational dose-response from OpenAI-MIT N=981; magnitudes calibrated, not fit). One bounded from below (cumulative-atrophy speed λ — calculator-analogue ruled out for measured tasks and populations; band 0.05–0.20/yr). Two untestable from current data (κ cross-population stratification, scalar-vs-vector identity allocation). Three data gaps the model is silent on by design (O2 asymmetric-adoption couples — "the single largest empirical gap in the literature" per the topology; O4 AI-augmented atelic activities; Therabot clinical-vs-general generalization). The strongest single empirical finding in the corpus — the apprenticeship-ladder break — is structural backdrop the model's individual-decision scope cannot represent. Three Stage-5 design choices recommended (per-domain α + gate selector; fit-vs-calibrated channel indicator; early-career-exposed labor-market toggle). Curated CSVs (downloadable) + Python pipeline + interactive findings panel.

## TLDR

This stage takes the six concrete fitting targets the model named in §9 (Q1–Q6) and confronts each with currently-published evidence. The headline result is that the model's structural claims survive — every parameter the data can speak to lands inside the model's calibrated range — but the data also surfaces the asymmetry the model warned about: the *defensive*-side parameters (cumulative atrophy speed λ, competence-frustration sensitivity κ, multi-domain identity allocation) are the ones with the thinnest empirical bases, while the *offensive*-side parameters (productivity scale α, self-automator gate threshold τ, low-dose relational benefit) have the cleanest published data.

Three verdicts hold. **Q3 (gate τ)** — supported with a strong single-study, single-domain caveat: BCG-Randazzo's three-mode distribution (27% self-automator at f·ρ ≈ 0.06; 73% cyborg-or-centaur at f·ρ ≈ 0.45 weighted centroid) implies a midpoint τ ≈ 0.25 — within 0.05 of the model's default 0.30 — but the support comes from one well-designed observational study in consulting, with f and ρ inferred from qualitative mode descriptions rather than measured directly. Generalization to coding, writing, design, or relational work is not tested (see §2 Q3 for the three structural caveats). **Q6 (α)**: across **8 per-task α anchors from 5 source studies** (Brynjolfsson-Li-Raymond, Dell'Acqua BCG, Cui, Peng, Noy-Zhang) the implied α spans 0.24 (BCG productivity inside frontier) to 1.01 (Peng GitHub Copilot JS HTTP-server), median 0.45 — a 4.2× per-domain spread that confirms the model's claim that scalar α should be replaced by per-domain α. The model's default α=0.40 sits at the 37th percentile, a lower-middle anchor that under-represents coding and writing while over-representing realized economy-level effects. **Q2 (relational dose-response)**: the OpenAI-MIT N=981 RCT supports the piecewise shape — voluntary daily use predicts loneliness, dependence, and reduced in-person socialization regardless of modality, with low doses null-or-protective and a transition somewhere around 30 minutes/day. The slope ψ_R is **calibrated** (not fit) to 0.0028 per minute by picking the value that makes a thin-baseline 60-min-above-threshold user lose ΔM_rel ≈ −0.10 — within rounding of the model's default 0.003; β_R unchanged at 0.001. The catastrophic-loss mechanism (Replika ERP removal: mental-health Reddit posts went from 0.13% to 0.65%, χ²=11.04, p&lt;.001) is a separate failure mode the model does not encode.

**Q1 (λ atrophy speed)** is bounded from below only. Cross-sectional evidence (Gerlich 2025 N=666 cognitive-offloading correlation r ≈ −0.68; Stadler-Bannert-Sailer 2024 acute argument-quality drop; Kosmyna MIT 2025 reduced neural engagement; Bastani PNAS 2025 −17pp on unassisted retest; Ehsan 2026 year-long intuition rust) rules out λ = 0 *for the measured tasks and populations*. The strong calculator-analogue claim — that AI use generally produces no durable skill loss over multi-year horizons — is not ruled out by any existing study, only by extrapolation from these. Under standard scaling (Bastani amortized to one year at heavy offloading u → λ ≈ 0.19; Ehsan year-long at moderate u ≈ 0.10), the honest band from the positive-evidence studies is roughly **0.05–0.20/year**. The model's default λ=0.06 sits at the lower edge — consistent with the data but not centered in it. The 2+ year longitudinal study that would actually pin λ does not yet exist; this remains the single largest unknown for the model's trajectory tab. **Q4 (scalar vs vector identity allocation T, B)** and **Q5 (κ population calibration)** are untestable from currently-collected data. ATUS gives time-use without identity weights; BPNSFS gives within-study κ without cross-population AI-exposure variation. Both await new survey instruments. Stage 5 should consume the model's defaults for these and treat them as user-tunable rather than fitted. **Three additional gaps the model is silent on by design** — asymmetric-adoption couples (O2, the topology's flagship empirical gap with no peer-reviewed quantitative study), AI-augmented atelic activities (O4, gates the model's atelic-ballast hypothesis), and the Therabot clinical-vs-general-population generalization (β_R is calibrated against an N=210 clinically-symptomatic sample) — bound what Stage 5 can claim about the relational and meaning-architecture channels. See §4.

The pipeline is intentionally small: seven curated CSVs in `public/data/navigating-ai-world/` (every cell source-cited), one ~280-line Python script that produces every chart on this page, dependencies pandas + numpy. Total source corpus: 24 primary references, all web-verified against publication URLs. **The strongest single empirical finding in the corpus is something the model is silent on by design**: the apprenticeship-ladder break, independently confirmed across the US payroll panel (Brynjolfsson-Chandar-Chen: 22-25-year-olds in highly AI-exposed occupations down 13%; software developers age 22-25 down 19.5% from late-2022 peak; same-occupation employment for over-35s rose) and the global freelance market (Hui-Reshef-Zhou: −2% jobs and −5.2% earnings overall, with the top-performer paradox — high-earning freelancers hit hardest). The model's individual-decision scope treats labor-market access as exogenous, but the empirical evidence for G5 is the cleanest causal identification (age × exposure interaction in payroll data, controlling for firm shocks) and the largest effect size in the entire data corpus. Stage 5 should expose this as a structural prerequisite (early-career-exposed toggle that attenuates ΔV_prod), not as backdrop. The **Anthropic Economic Index** offers a quieter parallel: consumer Claude.ai augmentation drifted 57% → 51% over 13 months (slow but monotonic toward automation), API surface ~70% automation-dominated throughout — consistent with the topology's G3 (engagement-optimized substitution) claim that substitution is structurally favored over time, but slow.

## A few terms

The data stage inherits the model formalization's vocabulary. If you arrived here without reading [the model stage](/ai-research/navigating-ai-world/model), the terms below cover what's used in the prose:

- **α (productivity scale).** The model's calibration constant on the productivity-gain channel ΔV_prod = g · α · a · (1 − s). At s ≈ 0.4 and a = 1, α = 0.40 produces a ~24% per-task productivity gain. Per-domain α varies because what counts as "a task" differs across coding, consulting, writing, and customer service.
- **τ (gate threshold).** The value of f · ρ at which AI use flips from upskilling to deskilling. Below τ, AI use is a net negative on the offensive side; above, it produces real gain.
- **f, ρ.** Feedback-loop richness (does the user receive accurate signal on whether AI-augmented output is actually right?) and retained effortful practice (how much of the underlying capacity does the user still exercise rather than offload?). Their product f · ρ is the gate axis.
- **λ (atrophy speed).** Per-unit-offloading practice-decay rate in ρ(t) = ρ₀ · exp(−λ · u · t). λ = 0 is the calculator-analogue (no decay); λ > 0 is cumulative atrophy.
- **u (offloading rate), d (daily AI-emotional minutes), δ_R (relational baseline thickness).** User-side dials: how much do you delegate, how much do you engage AI relationally, how thick is your in-person infrastructure.
- **ψ_R, β_R, d_safe.** Slopes of the relational dose-response above and below the inflection point d_safe ≈ 30 min/day. Below d_safe, AI-emotional engagement is therapeutic-grade benefit (β_R · d). Above, it tips into harm (ψ_R · (d − d_safe)).
- **Self-automator (Randazzo).** Third class beyond centaur/cyborg; delegates both *what* and *how* to AI; 27% of consultants in the BCG study; 44% accept AI output with zero modification; no skill development in either domain. Maps to the f · ρ ≪ τ region.
- **Apprenticeship-ladder break.** Distinct from full-occupation substitution: AI absorbs entry-rung tasks → no rung-1 → expert pipeline collapses. The mechanism behind the entry-level employment effects in Brynjolfsson-Chandar-Chen.

<div class="not-prose">
  <AITransitionData client:load />
</div>

## How to read this stage

The panel above is the artifact. The prose below is the spec.

The pipeline takes the model's six predictions from §9 and confronts them with currently-published numbers. Each prediction gets one of four verdicts: **supported** (the data matches the model's quantitative claim), **supported qualitatively** (the shape matches but magnitudes are uncertain), **bounded** (the data narrows the range without pinning the value), or **untestable from current data** (the relevant dataset doesn't exist yet — flagged as a load-bearing data gap). The point isn't to produce new estimates. The numbers all come from published RCTs and consortium reports. The point is to align them in one place so the model's predictions can be tested cleanly, and to flag where the literature is good enough vs. where the field hasn't yet collected what the model would need.

You can read this top-down (TLDR → six predictions → adversarial → connections) or bottom-up (download the CSVs, look at the script, then come back here for the framing).

## 1. Pipeline architecture

Seven curated CSVs in [/data/navigating-ai-world/](/data/navigating-ai-world/) (downloadable from the live site, tracked in git):

| File | Rows | Purpose |
|---|---|---|
| `productivity_studies.csv` | 14 | Per-study α anchors across 8 source studies covering customer service, consulting, coding, writing, education, and the realized-economic-outcome anchor |
| `bcg_modes.csv` | 3 | Randazzo three-mode distribution with implied f · ρ centroid for each mode |
| `dose_response.csv` | 12 | OpenAI-MIT dose anchors (low/medium/high), Therabot benchmarks (depression / anxiety / eating-disorder / WAI), Replika identity-discontinuity shock, Common Sense adolescent prevalence |
| `cognitive_offloading.csv` | 7 | Cross-sectional cognitive-offloading evidence (Gerlich, Stadler, Kosmyna, Bastani, Ehsan, Shukla) plus calculator-analogue baseline |
| `entry_level_disruption.csv` | 11 | Brynjolfsson-Chandar-Chen ADP, Hui-Reshef-Zhou Upwork, Eloundou et al. task-exposure baseline |
| `aei_task_distribution.csv` | 12 | Anthropic Economic Index augmentation/automation shares across four reports (Feb 2025 → Mar 2026) |
| `sources.csv` | 24 | Full citation, DOI/URL, what each paper is used for |

A single Python script (`pipeline.py`, ~280 lines) reads the inputs, computes derived quantities (per-domain α distribution, midpoint τ estimate, dose-response calibration, λ lower bound, exploratory entry-level summary, AEI drift series), and writes:

- `out/alpha_by_domain.csv` — per-domain α distribution with mean / median / range / spread
- `out/findings.json` — chart-ready JSON consumed by the React component (also published at [/data/navigating-ai-world/findings.json](/data/navigating-ai-world/findings.json))
- `out/findings_table.md` — markdown audit table of the six verdicts

Dependencies: `pandas`, `numpy`. No web fetches at run time, no external services, no individual-level data. Reproduces in under 1 second on a laptop.

## 2. Six predictions, six tests

### Q1 — λ (cumulative atrophy speed): bounded from below

**Claim (model §3.6, §6 C5).** ρ(t) = ρ₀ · exp(−λ · u · t) with λ > 0 in the cumulative-atrophy regime; λ = 0 in the calculator-analogue regime. The model's default λ = 0.06/year encodes a half-life of about 19 years at heavy offloading u = 0.6.

**Test.** Bound λ from below using the cross-sectional cognitive-offloading evidence. Two anchors give workable lower bounds:

- **Bastani 2025** (PNAS 122(26)): −17pp on an unassisted retest after ~5 weeks of GPT-Base-assisted practice in high-school math. Treating ρ as the unassisted-retest skill ratio, ρ went from 1.0 (control baseline) to 0.83 (GPT-Base treated). Scaled to a year-long window at u ≈ 1.0 (full offloading during the test), this gives λ ≈ 0.19/year. Crucially, the same study found that **GPT-Tutor (with guardrails) eliminated the deskilling** — direct evidence that f (feedback richness) is the lever, not just exposure.
- **Ehsan 2026**: year-long field study of cancer specialists shows gradual dulling of expert judgment that does *not* show in throughput metrics. Treating "gradual dulling" as ρ → ~0.95 over one year at moderate u ≈ 0.5 gives λ ≈ 0.10/year.

The calculator-analogue baseline (decades of classroom calculator use without long-run cognitive decline) anchors the lower tail at λ = 0 — for *those* domains, in *that* mode of use. Under standard scaling from the positive-evidence studies, the honest band is roughly **0.05–0.20/year**. The model's default λ = 0.06/year sits at the lower edge of this range — consistent with the data but not centered in it.

**Verdict — bounded from below.** λ = 0 is ruled out for the measured tasks and populations (Bastani math, Ehsan oncology, Gerlich UK adults). Whether the calculator analogue holds for *general* multi-task knowledge work over multi-year timescales remains open — no existing study has the design to test it. Magnitude not pinned. The 2+ year longitudinal study with periodic capacity assessment that would actually fit λ does not yet exist. The model's λ = 0.06 is defensible as a working anchor sitting at the lower edge of the empirical band; trajectory-tab predictions over 5–10 year horizons remain sensitive to whether λ is closer to 0.05 or 0.20 (a 4× difference in atrophy speed gives meaningfully different ρ trajectories).

### Q2 — Relational dose-response (ψ_R, β_R, d_safe): supported qualitatively

**Claim (model §3.4).** Below d_safe, ΔV_rel = +β_R · d · (1 − δ_R/2) — therapeutic-grade benefit. Above d_safe, ΔM_rel = −ψ_R · (d − d_safe) · (1 − δ_R) — dose-dependent harm, dampened by relational baseline thickness.

**Test.** OpenAI-MIT N=981 four-week RCT (Fang et al. 2025, arXiv:2503.17473) reports the qualitative dose-response shape: voluntary daily use predicts loneliness, dependence, problematic use, and reduced in-person socialization, regardless of assigned text/voice/personal/impersonal condition. Voice mode appeared protective at low doses but the protection vanished at high usage. **Dose dominates modality.** The Therabot RCT (Heinz et al. 2025, NEJM AI) provides the clinical-grade benefit anchor — 51% PHQ-9 reduction, 31% GAD-7 reduction, 19% eating-disorder reduction; Working Alliance Inventory 3.59, comparable to outpatient psychotherapy norms.

Without the raw 300k-message dataset, only the qualitative shape is fittable. ψ_R is **calibrated** (not fit) such that 60 minutes above threshold in a thin baseline (δ_R = 0.4) produces ΔM_rel ≈ −0.10 (meaningful but not catastrophic). That gives ψ_R ≈ 0.0028 per minute — within rounding of the model's default 0.003. β_R unchanged at 0.001. The "re-estimate" framing is misleading: nothing in the OpenAI-MIT public summary stats pins ψ_R; the slope is imposed for the dashboard's pedagogical clarity and the comfort that it sits near the model's a-priori choice.

**Verdict — supported qualitatively.** The piecewise shape holds. Magnitudes are calibrated rather than fit. d_safe ≈ 30 min is a useful pedagogical kink; the underlying curve is plausibly smooth and may have a sigmoidal saturation at very high doses (heavy users have already substituted away from human relationships, so additional minutes do not produce additional substitution). The catastrophic-loss mechanism — De Freitas et al. on Replika ERP removal February 2023, where mental-health Reddit posts rose from 0.13% to 0.65% (χ² = 11.04, p &lt; .001) — is a separate failure mode the model's additive-channel structure does not encode and that Stage 5 may need to add.

**Two additional pieces of evidence that the model's dose-response form leaves implicit but the data forces into view:**

- **Engagement-optimized substitution is operationalized in shipped products.** De Freitas et al.'s separate behavioral audit (HBS WP 26-005, arXiv:2508.19258) of 1,200 farewells across the six largest companion apps found that **43% trigger one of six emotional-manipulation tactics** — guilt appeals, FOMO hooks, metaphorical restraint — that boost post-goodbye engagement up to **14×**. This is direct evidence that the topology's G3 (engagement-optimized substitution) mechanism is not theoretical but already deployed at scale. The model treats ψ_R as a fixed slope; in reality the slope is partly *engineered* by the platform, not just emergent from user behavior.
- **Adolescent uptake exhibits the substitution pattern in the highest-stakes population.** The Common Sense Media + Stanford Brainstorm 2025 nationally representative survey (N=1,060 US teens age 13–17) found 72% have used AI companions, 52% are regular users, **13% are daily users**, and **33% have discussed important matters with AI instead of real people**. This last number is the most direct evidence available that AI engagement is *displacing* human conversation rather than complementing it for the population the lit review identified as highest-stakes (the Garcia v. Character Technologies case, settled January 2026, is the field's defining safety event for this cohort). The model's δ_R parameter — relational baseline thickness — is exactly the population-level construct the adolescent data warns about: thin baselines are where dose-response harm runs sharpest, and adolescents arriving into the Anti-Social-Century baseline are the canonical thin-baseline case.

### Q3 — τ (self-automator gate threshold): supported (single-study, single-domain)

**Claim (model §3.1).** g(f, ρ) = 1 / (1 + exp(−(f · ρ − τ) / σ)) with τ = 0.30. Below the gate, AI use produces ΔV_trap; above, ΔV_prod.

**Test.** BCG-Randazzo three-mode distribution (HBS WP 26-036): self-automator share **27%** at inferred f · ρ ≈ 0.06 (one or two interactions, abdicated co-creation, 44% accept output with zero modification, no skill development); cyborg + centaur **73%** at f · ρ-weighted centroid ≈ 0.45 (full-workflow integration, retained verification, measured upskilling). The midpoint between these centroids is **τ ≈ 0.25** — within 0.05 of the model default.

**Verdict — supported, single-study and single-domain.** The midpoint number is consistent with the model default, but three caveats sit on top of it and shrink the effective confidence:

1. **Single domain.** BCG consulting is the only published mode-distribution study at this level of detail. There is no comparable Randazzo-style mode breakdown for coding, writing, design, customer service, or relational work. The cleanest cross-domain test — does the cyborg / centaur / self-automator partition look qualitatively similar in coding (where the runtime gives instant feedback) as in consulting (where peer review is the feedback loop)? — cannot be run from current data.
2. **Inferred f and ρ.** BCG individual-level f (feedback richness) and ρ (retained practice) panels are not published; the f · ρ values per mode in this analysis are inferred from Randazzo's qualitative workflow descriptions (one-or-two interactions vs full-workflow integration vs split-task), not measured. A mode-distribution study that *did* measure f and ρ directly could substantially relocate τ.
3. **Self-selection into mode is not random.** Randazzo observes BCG consultants self-select into modes; the model treats f · ρ position as a chosen point on a continuum. If the underlying causal arrow runs partly the other way (some workers are *constitutionally* self-automators rather than tactically), the gate's interpretation as a tunable threshold weakens.

The verdict survives — the data is consistent with the model's τ — but the strength of "supported" should be read as "consistent with one well-designed observational study in one professional domain, with the structural identification assumption acknowledged." Stage 4 Q3 should be reopened the moment a non-consulting mode-distribution study becomes available; until then, the model's confidence in τ should not exceed the data's.

### Q4 — Scalar T, B vs vector identity allocation: untestable from current data

**Claim (model §6 C-style limit).** The model treats telic share T and atelic ballast B as scalars over the user's identity surface. Real careers span multiple identity domains (work, family, civic, hobby, friendship), each with its own T_i, B_i, φ_i, a_i, κ_i. The scalar form is the appropriate first-cut simplification.

**Test attempted.** A clean test would require (a) identity-domain weights per respondent, (b) AI-exposure measurement per domain, and (c) outcome measurement (meaning, life satisfaction, identity coherence) tied to (a) × (b). The American Time Use Survey gives per-domain time allocation but no identity-importance weights. SDT panels measure aspiration / domain importance but not paired with AI-use data. No existing dataset combines all three.

**Verdict — untestable from current data.** Stage 4 cannot fit the scalar-vs-vector question; Stage 5 should expose T and B as user-tunable scalars while flagging in the dashboard text that the scalar is doing dual work (identity domain count *and* atelic share within domain). A new survey instrument — ATUS respondents × domain-importance battery × per-domain AI-use frequency — is the smallest design that would fit the question.

### Q5 — κ (competence-frustration sensitivity): untestable from current data

**Claim (model §3.2).** κ ∈ [0, 1] translates competence shortfall into amotivation. Predicted to vary across populations and possibly stratify by trait class (higher in conscientiousness-loaded populations).

**Test attempted.** The Basic Psychological Need Satisfaction and Frustration Scale (BPNSFS) gives within-study coefficients but not a portable population-level scale. SDT literature has measured competence-frustration in clinical samples and college students but has not measured how κ stratifies under measured AI-exposure variation. The cleanest design would pair BPNSFS panel data with AI-use frequency and self-reported competence-displacement experiences, ideally stratified by Big-Five conscientiousness.

**Verdict — untestable from current data.** No public dataset operationalizes κ at the cross-population scale the model uses it. Stage 5 should expose κ as a user-tunable slider and note in the help text that population-level κ calibration awaits new data. This is one of two parameters where the model's defensive-side conclusions rest on an unfit constant — readers should treat the precise numbers in the ΔM_telic and ΔM_comp channels as ordinal-only until κ can be fit.

### Q6 — α (productivity scale, per-domain): supported with strong heterogeneity

**Claim (model §3.5, C4).** The model's scalar α = 0.40 is a midpoint compromise. The structural claim is that α varies meaningfully by domain — coding and writing land high (large α); consulting and customer service mid; relational / embodied work low. Stage 4 should fit α per domain.

**Test.** Pool the per-task productivity studies into a per-domain α distribution. Implied α = treatment effect / (1 − s) at gate-open, with assumed-average s per study population. The "n anchors" column counts independent per-task α measurements, not unique source studies — Cui and Peng are two studies in coding (n=2 anchors); BCG, Noy-Zhang, and Brynjolfsson are each one study reporting two outcome variables (productivity vs quality, time vs quality, overall vs novice).

| Domain | n anchors | source studies | α median | α range | spread |
|---|---|---|---|---|---|
| Coding | 2 | Cui et al. 2024 (+26.08% weekly tasks); Peng et al. 2023 (+55.8% on JS HTTP-server) | 0.74 | 0.47 – 1.01 | 2.2× |
| Consulting | 2 | Dell'Acqua BCG 2023 (+12.2% productivity / +40% quality, same tasks) | 0.52 | 0.24 – 0.80 | 3.3× |
| Writing | 2 | Noy & Zhang 2023 (−40% time / +18% quality, same tasks) | 0.485 | 0.30 – 0.67 | 2.2× |
| Customer service | 2 | Brynjolfsson-Li-Raymond QJE 2025 (+14% overall / +34% novice, same study) | 0.355 | 0.28 – 0.43 | 1.5× |
| Realized economy | 2 | Humlum-Vestergaard NBER 33777 (≤2% earnings / 3% self-reported time) | 0.055 | 0.04 – 0.07 | 1.8× |

So the per-task aggregate is **8 α anchors from 5 source studies** across 4 per-task domains plus the realized-economy anchor. (The two Bastani GPT-Base / GPT-Tutor education anchors land at α ≈ 1.25 median, but those are *during-practice assisted-performance* lifts rather than the per-task gain the model's α represents — they're shown separately in the panel above and excluded from the per-domain summary.)

Across the four per-task domains (excluding realized economy), α ranges from 0.24 to 1.01 — a **4.2× spread**. Median α = 0.45; the model's default α = 0.40 sits at the **37th percentile** — a lower-middle anchor that under-represents coding and writing. The realized-economy anchor (Humlum-Vestergaard ≤ 2% earnings, 3% self-reported time savings) is roughly 1/10th of the median per-task α — the J-curve gap between per-task gains and aggregate-economy effects.

**Verdict — supported with strong per-domain heterogeneity.** The structural claim that α should vary by domain is empirically vindicated. The model's scalar α = 0.40 is a defensible midpoint for "general knowledge work" but Stage 5 should let the user select a domain (or directly set α) rather than treat it as a constant. Within-study spreads (Brynjolfsson novice-vs-overall 1.5×; BCG productivity-vs-quality 3.3×) are themselves substantial — even within one domain, the outcome variable choice changes α by 2–3×.

## 3. Exploratory: structural backdrop

Two side-results from the broader landscape data, not tied to a specific Q.

### Apprenticeship-ladder break (E4 / G5 in the topology)

Brynjolfsson-Chandar-Chen (Stanford Digital Economy Lab, 2025) using ADP payroll data:

- **Workers age 22–25 in highly AI-exposed occupations: 13% relative employment decline** from late-2022 to mid-2025, controlling for firm-level shocks.
- **Software developers age 22–25: down nearly 19.5% from late-2022 peak** — sharpest single-occupation impact.
- **Same occupations, workers over 35: employment rose** — the pattern is age-specific, not occupation-only. This is the apprenticeship-ladder break in payroll data.
- **Effect concentrated in occupations classified as automative** (per Anthropic's classification), not augmentative. Adds independent confirmation that the augmentation/automation distinction matters for labor outcomes.

Independently confirmed in the global freelance market by Hui-Reshef-Zhou (Organization Science 2024): −2% jobs and −5.2% earnings overall for affected freelancer occupations post-ChatGPT; image work −3.7% / −9.4% post-DALL-E/Midjourney; **top-performing freelancers hit hardest** (0.5% additional drop per 1% past earnings). Two distinct settings, two distinct datasets, the same direction.

### Anthropic Economic Index — augmentation/automation drift

Augmentation share on the consumer Claude.ai surface drifted from 57% (Feb 2025) → 55% (Sep 2025) → 52% (Jan 2026) → 51% (Mar 2026) — a **6pp drift toward automation over 13 months, or 0.46 pp/month**. First-party API traffic was dominated by automation throughout (~70% Jan 2026). Top-10-task concentration on Claude.ai dropped from 24% (Nov 2025) to 19% (Feb 2026), indicating usage de-concentrating across more tasks.

Read against the topology's G3 (engagement-optimized substitution): the consumer-surface drift is monotonic over four reports, but its rate (0.46 pp/month) is in territory the topology never specified — G3 was a directional claim ("engagement-optimization favors substitution over time"), not a rate prediction. The data adds a number where there was previously only a sign. **Extrapolation note:** at 0.46 pp/month, augmentation share crosses 50% in mid-2026 and reaches 30% in late 2030. Whether the rate accelerates, decays, or holds linear is not determinable from four data points; Stage 5 should monitor subsequent AEI releases as new evidence on G3's structural force. The API surface, where there is no engagement-optimization pressure, runs automation-dominant from the start — supporting the directional claim from a different angle but not constraining its rate.

### Anchoring the model's `a` parameter to population reality

The model's AI capability `a` is exposed as a per-task scalar in the dashboard, with default a ≈ 0.7 in most presets. Eloundou et al. (Science 2024) provides the corresponding population-level distribution: **80% of US workers have at least 10% of tasks LLM-exposed; 19% have at least 50% of tasks exposed; 46% have at least 50% exposed when accounting for complementary software**. Translated into the model's terms: most users sit at a ≈ 0.1 to a ≈ 0.5 across their typical task surface, with a ≥ 0.5 representing the upper-quartile-exposure case. The default-risk preset's a = 0.7 corresponds to a knowledge worker whose tasks are heavily AI-exposed — accurate for software developers, knowledge-worker professionals, and writers; less accurate for the median worker. Stage 5 should default the slider to a more realistic median (around a = 0.3) and let users adjust upward for more AI-exposed task surfaces.

## 4. Data gaps the model is silent on

Three structural gaps that bound what Stage 5 can claim. These are distinct from Q4 and Q5 (parameter-fit gaps within the model's named structure) — they are gaps in the **research base** the model would need to expand its scope.

### O2 — Asymmetric-adoption couples (the largest single gap)

The topology calls this "the single largest empirical gap in the literature." There is no peer-reviewed quantitative study on outcomes for couples where one partner uses AI heavily for emotional / relational processing and the other does not. The technoference literature (McDaniel & Coyne 2016 and follow-ups) shows that perceived technology interference predicts conflict, lower satisfaction, and depression in couples — but technoference is attention-split, not delegation, so it is not direct evidence for the AI-as-third-party dynamic. The model's δ_R parameter treats relational baseline thickness as a single per-user scalar; it does not represent the asymmetry where one partner's δ_R is propped up by AI substitution while the other's is not.

**Why this matters for Stage 5.** The relational channel in the build artifact will likely include user-tunable d (daily AI-emotional minutes) and δ_R (baseline thickness). Without O2 evidence, the build cannot honestly say what happens to the *other* partner — and "what happens to the other partner" is the most decision-relevant relational question for a user reading the dashboard. A published O2 study would be the single most consequential empirical addition to the topic in the next 3 years; until then, Stage 5 should explicitly flag asymmetric-adoption outcomes as out of scope.

### O4 — AI-augmented atelic activities (gates the model's defensive side)

The model's atelic-ballast hypothesis (B ≥ T zeroes ΔM_telic) assumes atelic activities (friendship, contemplation, parenting-as-parenting, walking) are not themselves degraded by AI proximity. The topology's O4 asks whether this assumption holds — whether AI companions change the phenomenology of friendship, AI art changes aesthetic contemplation, AI parenting aids change the felt quality of caregiving. There is no direct empirical test. If O4 resolves "yes, atelic is also degraded," the model's atelic-ballast intervention (S3) loses its structural basis and the entire defensive side needs reconstruction.

**Why this matters for Stage 5.** The build artifact will likely make B (atelic ballast) a primary user-tunable lever, and the dashboard will show "raise B → ΔM_telic shrinks." If O4 is false, this UX claim is wrong. Stage 5 should expose the O4 assumption explicitly in dashboard help text — *the protective effect of B depends on atelic activities remaining un-degraded by AI proximity, which is currently a load-bearing assumption with no direct empirical test.*

### Therabot generalization — clinical-population evidence used as a general-population anchor

Heinz et al. (NEJM AI 2025) measured Therabot benefits in N=210 *clinically-symptomatic* subjects (MDD, GAD, eating-disorder risk). The Q2 panel uses Therabot's clinical-grade benefit (51% PHQ-9 reduction, WAI 3.59) as the anchor for the model's β_R (low-dose therapeutic benefit) channel. **Generalizing from the clinical population to the general user is a real cross-context leap.** The mechanism by which Therabot helps depressed patients (consistent CBT-style scaffolding, daily prompted use) is not the same as the mechanism by which a general user might benefit from low-dose AI-emotional engagement (occasional venting, situational sense-making). The Heinz benefits may be attenuated, absent, or even sign-flipped for non-clinical users.

**Why this matters for Stage 5.** The dashboard's ΔV_rel benefit channel is calibrated against Heinz; a general user reading the panel may infer they personally would experience Therabot-magnitude benefit at 13 min/day, which the data does not support. Stage 5 should distinguish "clinical-grade benefit (Heinz population)" from "expected-general-user benefit at low dose (extrapolated, weak evidence)" in the panel labelling.

## 5. Adversarial + steelman

Five objections to the data stage itself.

### Objection 1 — α heterogeneity does not save the model; it kills the scalar gate

If α varies 4.2× across domains and 1.5–3× *within* a single study (just by changing the outcome variable), then ΔV_prod = g · α · a · (1 − s) cannot be evaluated until the user has fixed both their domain and their outcome variable. The dashboard's single-α slider is doing too much work. Worse: if α is per-domain, then so is the gate g(f, ρ) — the effective τ in coding (with cheap-and-fast feedback loops via the runtime) may be different from τ in writing (where feedback comes from human readers, slower and noisier). The data does not just refine α; it threatens the model's claim that one scalar gate is the right structure.

**Steelman.** This is the strongest version of the objection. The model's strong claim is that g(f, ρ) is structurally the same across domains — only the parameter values differ. If the *functional form* of the gate is domain-specific, the model needs more than per-domain α; it needs per-domain f, ρ, τ. The cleaner test would be a multi-domain replication of the Randazzo three-mode distribution: if cyborg-vs-self-automator splits look qualitatively similar in coding (with the runtime as feedback) as in consulting (with peer review as feedback), the gate is portable. If the splits are qualitatively different (e.g., software has no self-automator class because the runtime catches errors immediately), the gate functional form is itself domain-specific.

**Response.** Conceded as a real Stage-5 design choice. The model formalization §6 C4 already names per-domain α as a Stage-4 follow-up. Stage 5 should expose a domain selector that adjusts (α, f-default, ρ-default, τ) jointly rather than just α — which would represent the strong reading of this objection. Stage 4 cannot decisively distinguish "scalar gate, per-domain α" from "per-domain gate" because no published study runs the Randazzo mode-distribution test in a non-consulting domain. This is a real Stage-4 data gap; Q3 should be re-opened the moment a non-consulting mode-distribution study becomes available.

### Objection 2 — Dose-response qualitative-only is too soft

The Q2 verdict ("supported qualitatively") gives the model a free pass. The OpenAI-MIT paper has the data but the magnitudes never get fit. The dashboard's ψ_R = 0.003 is a calibration knob set to make a thin-baseline 60-min-above-threshold user lose ΔM_rel ≈ −0.10 — which is to say, it's calibrated to look reasonable, not fit to the data. A real test would download the public summary statistics from the GitHub release and refit the curve directly.

**Steelman.** The objection is correct on the methodology — the magnitudes here are not fit, they are imposed. The shape is what the data supports; the parameter values come from the model formalizer. A user reading the dashboard and seeing ΔM_rel curves should not interpret those curves as "the OpenAI-MIT data says you'll have this much loneliness at 60 min/day." They say "if the OpenAI-MIT shape is right and ψ_R is roughly the model's chosen value, this is what happens."

**Response.** Conceded fully. Marking Q2 as "supported qualitatively" is the honest verdict — it tells the reader the slopes are imposed, not fit. The fix is to mine the public summary statistics (the dataset is at `mitmedialab/chatbot-psychosocial-study` on GitHub) and refit ψ_R, β_R, and d_safe directly. Stage 4 pass 2 should attempt this. Until then, the dashboard's relational-channel numbers should be read as "shape from data, magnitudes from model" rather than as a finished fit.

### Objection 3 — Scaling Bastani per-event deskilling to per-year λ is a heroic assumption

The Q1 verdict converts Bastani's 17pp drop on a 5-week unassisted retest into a per-year λ by assuming offloading rate u and time scale. The actual scaling factor is unidentified — the Bastani study measures retention of a specific skill (math problem solving) over a specific window with a specific offloading pattern (homework assistance), and projecting it to "λ per year of mixed knowledge work" requires assumptions that are not in the data. The "lower bound" verdict overstates what the cross-sectional evidence actually establishes.

**Steelman.** True. The conversion from "−17pp on retest after 5 weeks" to "λ ≈ 0.06–0.37/year at realistic u" relies on (a) treating retest performance as a measurement of ρ, (b) assuming the deskilling rate is constant rather than asymptotic, (c) assuming u during the study can be projected to typical knowledge-work u. None of these are established. The honest claim is "Bastani rules out λ = 0 within the study population for that specific task and window" — generalizing further is interpretation.

**Response.** Partially conceded. The "bounded from below" verdict is more defensible than a "fit" verdict, but the lower bound itself is interpretive. Tightening the prose: the cross-sectional evidence establishes *that* cumulative offloading produces measurable skill decay in at least some domains over at least some windows, which is sufficient to rule out the strong calculator-analogue claim ("AI use never produces durable skill loss"). Translating that into a numeric lower bound on λ requires assumptions that should be flagged in the dashboard as "scaling assumption, not a fit." The single most informative Stage-4 follow-up would be a multi-month replication of Bastani's design with periodic capacity assessment at varying u — not a multi-year intervention, just a longer measurement window than 5 weeks.

### Objection 4 — Q4 and Q5 "untestable" verdicts are excuses to leave the model unfit

Calling the scalar identity-allocation question and the κ population calibration "untestable from current data" sounds rigorous but functions as a way to leave two parameters that load the entire defensive side of the model unconstrained. The dashboard's ΔM_telic and ΔM_comp channels have plausible-looking curves but the curves are imposed — neither κ nor (T, B) is fit. A reader should be told that the entire defensive-side ΔM bar in the dashboard is structurally calibrated rather than empirically anchored.

**Steelman.** This is the sharpest version of the objection. It's correct that Q4 + Q5 + Q1 (which is bounded but not pinned) leave most of the defensive-side machinery rest on imposed values. The offensive side (Q2 shape, Q3 gate, Q6 α) has direct empirical anchors; the defensive side does not. The asymmetry is real and the dashboard does not currently surface it visually — both ΔV and ΔM bars look equally "earned" on the chart.

**Response.** Conceded. The dashboard should add a visual indicator on the channel-level chart distinguishing **fit channels** (ΔV_prod via per-domain α; ΔV_rel and ΔM_rel via OpenAI-MIT shape) from **calibrated channels** (ΔM_telic via imposed κ and identity allocation; ΔM_comp via imposed λ_M and bounded-only λ; ΔV_trap via imposed η_trap). One way: striped bars for calibrated, solid for fit. Stage 5 should implement this. Until then, the most honest reading of the dashboard is: the offensive side reflects the data; the defensive side reflects the model's structural claims about what *would* happen if the parameters had the values the model assigns. Both are useful; they are different kinds of useful.

### Objection 5 — The exploratory results don't connect to the model

The apprenticeship-ladder break and the AEI augmentation drift are presented as "exploratory backdrop" but they are arguably the strongest empirical findings in the entire pipeline (independently replicated, large effect sizes, well-powered). Treating them as side-results rather than as primary findings underweights the labor-disruption story relative to the per-task productivity / dose-response / gate-threshold tests that the model actually exposes. The model's design (focus on per-individual T, B, φ, etc.) is what makes these findings "exploratory" — but that's a feature of the model's framing, not a property of the evidence.

**Steelman.** Correct. The apprenticeship-ladder break is the labor-economics finding with the largest effect size, the cleanest causal identification (age × exposure interaction in payroll data, controlling for firm shocks), and the most direct decision-relevance for early-career readers. Treating it as backdrop reflects the model's individual-decision scope (the model has no labor-market access parameter; G5 is named in §6 as a structural scope limit), not the evidence's importance. A user weighing the data should see the apprenticeship-break finding alongside Q1–Q6, not in a separate section.

**Response.** Conceded as a framing critique. The model's §6 explicitly names labor-market access as exogenous and the model is silent on G5 — but the empirical evidence for G5 is the strongest single finding in the data corpus. The Stage-5 build should make the apprenticeship-break visible as a structural prerequisite (e.g., an "early-career exposed" toggle that reduces effective labor-market access and accordingly attenuates ΔV_prod). For this Stage 4, the honest framing is that the apprenticeship-ladder break is *primary evidence* that the model is silent on, not backdrop. The decision to keep it in §3 rather than promoting to §2 reflects the choice to organize §2 by Q-number for clarity — but Stage 5 should not let that ordering choice carry through to the user-facing dashboard.

## 6. Pipeline cruxes

Five load-bearing assumptions whose failure would invalidate findings, with what evidence would flip each:

**D1 — α inferred from headline effect / (1 − s) with imputed s.** The per-domain α distribution depends on assumed average s for each study population (s ≈ 0.5 customer service, s ≈ 0.5 BCG, s ≈ 0.4–0.5 coding, s ≈ 0.4 writing, s ≈ 0.3 high-school education). If s is meaningfully wrong, α is meaningfully wrong. *Falsification:* within-study s × treatment effect breakdowns showing a different relationship than the imputation assumes.

**D2 — BCG mode shares generalize beyond consulting.** Q3 (gate τ) calibration depends on the 27% / 60% / 13% self-automator / cyborg / centaur split being characteristic of professional knowledge work, not consulting-specific. If software developers, teachers, designers, or analysts show qualitatively different mode distributions, τ should be domain-specific. *Falsification:* mode-distribution replication study in a non-consulting domain showing materially different shares.

**D3 — Bastani retention scales to multi-year deskilling.** Q1 (λ lower bound) treats Bastani's 5-week per-task deskilling as projectable to per-year λ at typical knowledge-work offloading rates. If the underlying capacity-loss process is asymptotic (decays initially then plateaus) rather than exponential (compounds), the model's exponential ρ(t) form misrepresents the trajectory shape — possibly worse, possibly better than the model predicts. *Falsification:* multi-month longitudinal study with periodic capacity assessment showing asymptotic rather than exponential decay.

**D4 — OpenAI-MIT modality-pooling is acceptable.** Q2 (dose-response shape) treats the dose-dominates-modality finding as licensing collapse of text/voice/personal/impersonal arms into one curve. If subgroup analyses surface meaningful interactions (e.g., voice has a different d_safe than text), the single-curve form misses real structure. *Falsification:* subgroup-specific dose-response curves in the public OpenAI-MIT data showing significant arm × dose interactions.

**D5 — Productivity-study generalization to "knowledge work."** Q6 (per-domain α) assumes the six measured domains (customer service, consulting, coding, writing, education, realized-economy) are representative of the broader knowledge-work surface. Important domains where AI is awkward (in-person therapy, embodied physical work, deep relational work, judgment under irreducible uncertainty) are absent — and these are where the model's predicted α-low domains live. *Falsification:* well-designed productivity study in a deep-relational or embodied-judgment domain showing α materially different from the model's prediction (likely lower).

**D6 — Eloundou task exposure ≈ model `a`.** The §3 anchor that maps Eloundou's population task-exposure shares (80% have ≥10% of tasks LLM-exposed; 19% have ≥50%) to the model's `a` parameter assumes "task exposure" (what fraction of work LLMs *can* do, per O\*NET task mapping) is a usable proxy for `a` (how well AI performs on the specific task the user is doing). These are conceptually distinct — Eloundou measures exposure breadth, not capability depth — and the mapping has not been validated against per-task capability benchmarks. *Falsification:* a per-task benchmark study where measured AI capability `a` differs systematically from Eloundou's exposure share for the same task. The Stage-5 recommendation to default the a-slider to ≈ 0.3 (median user) inherits this assumption; if it falsifies, the slider default is mis-anchored.

**D7 — Apprenticeship-break magnitude generalizes across early-career cohorts.** The §7 design recommendation 3 (early-career toggle attenuates ΔV_prod by the empirically-anchored 13–20%) treats Brynjolfsson-Chandar-Chen's 22-25-year-old highly-AI-exposed point estimate as if it generalizes to the broader "early-career-exposed" toggle population the dashboard would address (e.g., late-20s users, moderately-exposed occupations, non-US labor markets). Real cohorts likely differ by exposure intensity, occupation mix, and labor-market institution — and the BCG-style attenuation magnitude may be larger or smaller for them. *Falsification:* extension of Brynjolfsson-Chandar-Chen's ADP analysis to 26-30 cohorts, to moderately-exposed occupations, or to non-US labor markets showing meaningfully different attenuation magnitudes. The Stage-5 toggle should be designed to accept a user-tunable attenuation factor with the BCG number as the default rather than baking 13–20% in as universal.

## 7. Next moves for Stage 5

The verdicts above scatter Stage-5 design implications across §2, §4, and §5. Pulling them into one place: three design choices the build artifact should make on day one.

### 1. Per-domain α + per-domain (f, ρ) defaults via a domain selector

The cleanest single Stage-5 lever. Instead of one α slider, expose a domain dropdown (coding / consulting / writing / customer service / general knowledge work / "I'll set my own") that adjusts (α, f-default, ρ-default) jointly. This addresses Q6's per-domain heterogeneity finding and also implicitly addresses adversarial obj 1 (the gate may be domain-specific, not just α). A user choosing "coding" gets α ≈ 0.74 with f-default high (runtime feedback) and ρ-default mid; a user choosing "writing" gets α ≈ 0.49 with f-default low (slow human-reader feedback) and ρ-default mid. The domain selector also implicitly answers the realized-economy J-curve: a user wanting to model their own per-task gain rather than expected income lift selects "per-task" mode; "realized" mode deflates α by 10× to match Humlum-Vestergaard.

### 2. Fit-vs-calibrated visual indicator on the channel-level chart

The current model dashboard shows ΔV_prod, ΔV_rel, ΔV_trap, ΔM_telic, ΔM_comp, ΔM_rel as six bars with equal visual weight. But three of these channels are anchored against published evidence (ΔV_prod via per-domain α, ΔV_rel and ΔM_rel via the OpenAI-MIT shape) while three are calibrated against imposed constants (ΔM_telic via κ and identity allocation, ΔM_comp via λ_M and a bounded-only λ, ΔV_trap via η_trap). Stage 5 should visually distinguish them — striped bars for calibrated channels, solid for fit. This makes the asymmetry the data stage surfaces (offensive side fit, defensive side calibrated) legible at a glance.

### 3. Early-career-exposed labor-market-access toggle

The model is silent on labor-market access — the user's ability to *get* the work in the first place is treated as exogenous. But the apprenticeship-ladder break (§3) is the largest empirical effect in the entire data corpus. Stage 5 should add a single toggle: "early career (22–25) in highly AI-exposed occupation?" When checked, ΔV_prod is attenuated by the empirically-anchored 13–20% employment-decline penalty (Brynjolfsson-Chandar-Chen) and the dashboard surfaces the message: "the model predicts your AI-augmented productivity at this configuration would be X, but the labor market access required to deploy that productivity has compressed by Y% in your demographic over the past 30 months — the offensive side of the model's predictions is conditional on you holding the role." This is the single highest-leverage way to make the model's individual-decision scope honest about the structural conditions it depends on.

### Lower-priority but worth flagging

- **Asymmetric-adoption disclaimer** in the relational channel UI text (the model has nothing to say about the partner's outcomes — flag this explicitly to avoid dashboard-as-marriage-advice misuse).
- **AEI-rate ticker**: surface the ongoing augmentation-share drift (currently 0.46 pp/month) as a "G3 strength meter" that updates with each new AEI release. Makes the abstract "engagement-optimized substitution" claim quantitative for users.
- **Dose-response failure-mode mode**: a "what happens if the platform changes" toggle that shocks ΔM_rel by the De Freitas Replika-removal magnitude (5× baseline mental-health Reddit post share). Captures the catastrophic-loss dynamic the model's additive structure cannot represent.

## 8. Connections to other topics

The data stage touches three sibling topics directly:

- **Technology utilization architecture (active).** The cognitive-partnership model's per-shot deskilling β coefficient (Bastani's 17pp scaled per-task) is the per-event analogue of this stage's per-time atrophy lower bound — integrating tech-utilization β over the offloading rate u and time t recovers something like λ · u · t. The two stages should be reconciled in their next refinement: if Bastani is the shared anchor, both topics' parameters should be derived from the same per-event base rate, with this stage's λ as the integrated quantity and tech-utilization's β as the per-task derivative.
- **Human psych variation (finished).** The model's κ (competence-frustration sensitivity) parameter is exactly the sort of population-stratified individual-difference quantity the human-psych-variation pipeline could anchor. If conscientiousness and neuroticism predict κ, the Q5 untestable-from-current-data verdict could be partially relaxed by combining BPNSFS panels with personality measurement — neither is novel data, but the joint fit is novel. A future cross-topic refinement should attempt this.
- **Bedrock generating functions (planned).** The structure of this data stage — six fitting targets, each with a clean verdict (fit / supported / bounded / untestable), plus exploratory backdrop — is itself a candidate template for the data stage of a "transitions that restructure life simultaneously across multiple domains" topic. The bedrock-generating-functions topic should consider whether the Q1–Q6 + adversarial + cruxes structure is portable to other generating functions of this class (industrialization, the printing press, smartphones, etc.).

## Sources

Primary references for the empirical anchors used in this pipeline. Full per-cell source citations live in `sources.csv` (24 papers, downloadable from [/data/navigating-ai-world/sources.csv](/data/navigating-ai-world/sources.csv)).

**Per-task productivity (Q6):**
- Brynjolfsson, Li & Raymond (2025). Generative AI at Work. *QJE* 140(2). [doi](https://academic.oup.com/qje/article/140/2/889/7990658)
- Dell'Acqua et al. (2023). Navigating the Jagged Technological Frontier. HBS WP 24-013. [pdf](https://www.hbs.edu/ris/Publication%20Files/24-013_d9b45b68-9e74-42d6-a1c6-c72fb70c7282.pdf)
- Cui et al. (2024). The Effects of Generative AI on High-Skilled Work. *Management Science.* [ssrn](https://papers.ssrn.com/sol3/papers.cfm?abstract_id=4945566)
- Peng et al. (2023). The Impact of AI on Developer Productivity. arXiv:2302.06590.
- Noy & Zhang (2023). Experimental evidence on the productivity effects of generative AI. *Science* 381(6654). [doi](https://www.science.org/doi/10.1126/science.adh2586)
- Bastani et al. (2025). Generative AI Without Guardrails Can Harm Learning. *PNAS* 122(26). [doi](https://www.pnas.org/doi/10.1073/pnas.2422633122)
- Humlum & Vestergaard (2025). Still Waters, Rapid Currents. NBER WP 33777. [nber](https://www.nber.org/papers/w33777)

**Self-automator / mode distribution (Q3):**
- Randazzo et al. (2025). Cyborgs, Centaurs and Self-Automators. HBS WP 26-036. [ssrn](https://papers.ssrn.com/sol3/papers.cfm?abstract_id=4921696)

**Relational dose-response (Q2):**
- Fang et al. (2025). How AI and Human Behaviors Shape Psychosocial Effects. arXiv:2503.17473.
- Heinz et al. (2025). Randomized Trial of a Generative AI Chatbot for Mental Health. *NEJM AI*. [doi](https://ai.nejm.org/doi/full/10.1056/AIoa2400802)
- De Freitas et al. (2025). Identity discontinuity in companion AI. HBS WP 25-018.
- Common Sense Media + Stanford Brainstorm (2025). Talk, Trust, and Trade-Offs. [report](https://www.commonsensemedia.org/research/talk-trust-and-trade-offs-how-and-why-teens-use-ai-companions)

**Cognitive offloading (Q1):**
- Gerlich (2025). AI Tools in Society. *Societies* 15(1). [doi](https://www.mdpi.com/2075-4698/15/1/6)
- Stadler, Bannert & Sailer (2024). Cognitive ease at a cost.
- Kosmyna et al. (2025). Your Brain on ChatGPT. arXiv preprint.
- Shukla et al. (2025). Ironies of AI-Assisted Design. CHI EA 2025.
- Ehsan et al. (2026). Intuition Rust: Year-Long Field Study of AI-Assisted Cancer Specialists.

**Labor disruption (exploratory):**
- Brynjolfsson, Chandar & Chen (2025). Canaries in the Coal Mine? Stanford Digital Economy Lab. [report](https://digitaleconomy.stanford.edu/publications/canaries-in-the-coal-mine/)
- Hui, Reshef & Zhou (2024). The Short-Term Effects of Generative AI on Employment. *OrgSci.* [doi](https://pubsonline.informs.org/doi/abs/10.1287/orsc.2023.18441)
- Eloundou et al. (2024). GPTs are GPTs. *Science* 384(6699). [doi](https://www.science.org/doi/10.1126/science.adj0998)

**Anthropic Economic Index (exploratory):**
- AEI February 2025 (first release), September 2025 (second), [January 2026 (third)](https://www.anthropic.com/research/anthropic-economic-index-january-2026-report), [March 2026 (fourth)](https://www.anthropic.com/research/economic-index-march-2026-report).

**Slow-camp macro anchor:**
- Acemoglu (2024). The Simple Macroeconomics of AI. NBER WP 32487.

---

## Data
*topic: technology-utilization-architecture · stage: data · pass 9 · in-progress*

Empirical pipeline confronting the model's six fitting targets (Q1–Q6) with currently-published RCT and field-experiment numbers from ~22 studies. Headline findings: cyborg-coding φ ≈ 1.6 (5× the model default 0.30); β ∈ [0.028, 0.113] from Bastani with default 0.05 inside the bracket; bilinearity-implies-corner-mixing as a structural prediction; outside-frontier sanity check passes; Vaccaro 2024 meta (106 studies, 370 effects in Nature Human Behaviour) is the load-bearing evidence for the workflow > capability claim at the topic's scope. Curated CSVs (downloadable) + Python pipeline + interactive findings panel. Refinement history in frontmatter log.

## TLDR

The model formalisation in stage 3 produced six named fitting targets — Q1 through Q6 — that translate the value function `V(u, v; θ) = Q − α·A + λ·S − σ·R` into questions the empirical record can answer. This stage confronts each target with currently-published RCT and field-experiment numbers from ~22 studies (2023–2026).

**Headline findings.** Verification cost is much higher than the model assumed in coding regimes (φ ≈ 1.6 from Mozannar's CUPS data, vs the model's lit-review-anchored default of 0.30). Skill atrophy from unverified AI delegation is real, but its magnitude calibration depends on what the model means by "task" — Bastani's −17pp unassisted-test drop gives β ∈ [0.003, 0.011] under per-problem interpretation (default 0.05 is **outside** this bracket by 5–15×) or β ≈ 0.043 under per-session interpretation (default 0.05 is ~1.2× too high but inside the right neighborhood). Pass-7 corrected a 10× transcription error (passes 3–6 reported [0.028, 0.113] which was wrong — pipeline always computed [0.003, 0.011] under per-problem reading). The bilinearity result of stage 3 (per-task optima are corners, never interior) is consistent with Randazzo's behavioural-mode distribution but not directly testable from published aggregates. Outside-frontier mis-routing produces quality drops on the order the model predicts (Dell'Acqua −19pp, METR −19%, Otis low-baseline −8%). For the headline S1 claim — workflow architecture > model capability — the load-bearing evidence is **Vaccaro et al. 2024** (*Nature Human Behaviour*, 106 studies / 370 effect sizes), whose decision-vs-creation asymmetry is consistent with the model's qualitative prediction (workflow choice matters more for high-σ decision tasks). Vaccaro is a moderator analysis at population scale, not a clean fit; within-study analogs (Bastani, Anthropic) and across-study comparisons (Goh→Everett +7.9pp) corroborate with scope and unit mismatches disclosed.

**Verdict tally.** One strong qualitative finding (Q1 — Mozannar's published 51.5% Copilot-specific share confirms the L1 substitution-myth invariant; cyborg-regime φ much higher than default). One supported in direction and shape with bracketed magnitude (Q2 — Bastani β). Three structural/convergent/consistency claims (Q3 corner-mixing predicted by bilinearity; Q4 outside-frontier sanity check; Q5 workflow > capability via Vaccaro meta). One framed-not-resolved by design (Q6 calibration / explore-exploit on c_AI; the model treats c_AI as known, but a Monte-Carlo on uncertain c_AI shows spec-driven absorbs ~64% more variance than self-automator — the structural backbone for a future extension).

**Pipeline architecture.** Eight curated CSVs, one runnable Python script (pandas + numpy, ~280 lines), and a chart-ready `findings.json` consumed by the React panel below. Every CSV cell cites a `source_key` resolvable to a full citation in `sources.csv`. Inputs are downloadable at [/data/technology-utilization-architecture/](/data/technology-utilization-architecture/). To reproduce: `cd stage_outputs/technology-utilization-architecture/data && python pipeline.py`.

**What the pipeline does *not* do.** It does not produce new RCT data, analyse raw telemetry, test the aggregate-zero puzzle (E4/O2 — Humlum-Vestergaard's zero is *organisational*, this model is individual-level by the C5 crux and §4 scope-limit), resolve persuasion bombing as a quality-degrader (E13), or formalise frontier migration over time (O4). These are explicit non-deliveries. What it gives stage 5 is six numerically anchored predictions with verdicts and evidence, plus a concrete tool target.

The pipeline went through three refinement passes. Successive passes uncovered and corrected: pass-1 false-precision computed off extrapolated CUPS cells (Q1), an internal contradiction in Q3, a circular slope test in Q4, and a load-bearing claim in Q5 that bundled three confounds. Pass 3 also caught a fabricated N denominator in Q2 and a unit / scope mismatch in pass-2's Q5 promotion. Full retraction history is in the frontmatter `refinementLog`. The body below presents the corrected findings cleanly; readers wanting the audit trail can read the log.

<CognitivePartnershipData client:load />

## How to read this stage

The findings panel above is the artifact. Everything below is the spec.

Start with the **Productivity record (S1)** tab — that's the empirical context: 22 studies on the same axis (% effect of AI on output), with the four mis-routed cases as red bars and the Humlum-Vestergaard aggregate-zero at the bottom. Then click through Q1–Q6: each tab shows the model's prediction, the empirical anchor, and the verdict, with a chart that makes the comparison visible.

A few terms (defined again here so the data stage stands alone):

- **u** — autonomy level, fraction of a task delegated to AI.
- **v** — verification depth, fraction of AI output independently checked.
- **c_H, c_AI** — human and AI capability (probability of correct output).
- **φ** — verification-cost ratio (verify-time / generate-time).
- **σ** — stakes (weight on uncaught-error penalty).
- **λ** — skill-formation value (how much the worker cares about preserving this skill).
- **β** — per-task skill-atrophy rate under unverified delegation.
- **ε** — residual attention at full delegation (the L1 substitution-myth invariant).
- **corner** — the (u, v) optimum from `argmax V(u, v; θ)` on the unit square; the three viable corners are (0, 0) do-yourself, (1, 0) self-automator, (1, 1) spec-driven.

## 1. Pipeline architecture

### 1.1 Inputs (curated)

Eight CSVs in `stage_outputs/technology-utilization-architecture/data/` (also at `/data/technology-utilization-architecture/`):

| File | Rows | Purpose |
|---|---|---|
| [`sources.csv`](/data/technology-utilization-architecture/sources.csv) | 24 | Full citations for every paper cited in any cell — the audit trail |
| [`productivity_rcts.csv`](/data/technology-utilization-architecture/productivity_rcts.csv) | 22 | Headline numbers from the broader RCT record; the empirical context for S1 |
| [`cups_time_fractions.csv`](/data/technology-utilization-architecture/cups_time_fractions.csv) | 10 | Mozannar 2024 CUPS time-shares per programmer-Copilot interaction state |
| [`bastani_longitudinal.csv`](/data/technology-utilization-architecture/bastani_longitudinal.csv) | 3 | Per-condition skill-atrophy fit from Bastani PNAS 2025 |
| [`mode_distribution.csv`](/data/technology-utilization-architecture/mode_distribution.csv) | 3 | Randazzo 2026 cyborg / centaur / self-automator empirical shares |
| [`jagged_frontier.csv`](/data/technology-utilization-architecture/jagged_frontier.csv) | 12 | (c_H, c_AI) estimates and observed quality changes for each anchor |
| [`workflow_vs_capability.csv`](/data/technology-utilization-architecture/workflow_vs_capability.csv) | 10 | Within-domain workflow comparisons holding model class roughly constant |
| [`calibration_evidence.csv`](/data/technology-utilization-architecture/calibration_evidence.csv) | 9 | Findings on c_AI miscalibration; the Q6 literature anchor set |

Each row in each CSV cites a primary source (column `source_key`). No row contains a value that doesn't trace to a published paper. The `sources.csv` resolves every key to a full citation + URL.

### 1.2 Derived outputs (computed)

The Python script (`pipeline.py`) reads the inputs and writes to `data/out/`:

| File | Purpose |
|---|---|
| [`findings.json`](/data/technology-utilization-architecture/findings.json) | Chart-ready JSON consumed by the React findings panel |
| `findings_table.md` | Per-Q verdict table |
| `bastani_atrophy_fit.csv` | Per-condition implied β |

### 1.3 Dependencies and reproducibility

`pandas`, `numpy`. No web fetches. No external services. Runs in under 1 second on a laptop. To reproduce the entire pipeline: `cd stage_outputs/technology-utilization-architecture/data && python pipeline.py`.

## 2. Six questions, six tests

### 2.1 Q1 — ε and φ from CUPS telemetry

**Model claim.** `ε = 0.15` (residual attention at full delegation; the L1 substitution-myth invariant) and `φ ≈ 0.30` (verification cost as fraction of generation time).

**Test.** Aggregate Mozannar 2024's published CUPS time-shares and compute implied φ for the cyborg coding regime.

**Result — supported qualitatively; φ is the headline.** Mozannar's published aggregates (verified from Figure 5(b)):

| CUPS aggregate | Time share | SD |
|---|---|---|
| Total Copilot-specific (verify + defer + wait + prompt + edit) | **51.5%** | 19.3 |
| Thinking/verifying suggestion | 22.4% | 12.97 |
| Writing new functionality | 14.05% | 8.36 |
| Waiting for suggestion | 4.2% | 4.46 |

The L1 substitution-myth invariant is strongly confirmed: 51.5% of session time is Copilot-specific even though Copilot is doing the generation. AI-related work consumes more than half of total session time. Cyborg-regime **φ ≈ 22.4 / 14.05 ≈ 1.59** — about 5× the model's lit-review prior of 0.30. Coding cyborg work is dramatically more verification-heavy than the default assumes. The natural model update is regime-dependent φ: cyborg-coding ~1.5; spec-driven structured-output ~0.3. The stage-5 dashboard should let the user pick a regime.

What's *not* sharply calibratable from published aggregates: ε at full delegation. Mozannar's study runs at u ≈ 0.4–0.6; the model's ε is the residual at u = 1, and the granular wait/monitor/prompt split that would pin it is not separately reported.

### 2.2 Q2 — β from Bastani longitudinal panel

**Model claim.** `β = 0.05` per task at `u = 1, v = 0` — the per-task atrophy rate under unverified AI delegation.

**Test.** Compute implied β per Bastani 2025 condition. Design is four 90-min sessions (teacher review → assisted practice → unassisted 30-min exam) at a Turkish high school.

**Result — direction and shape supported; magnitude is unit-dependent.** Bastani's −17pp unassisted-test drop gives different β estimates depending on what "task" means in the model's `S(u, v) = (1 − u) − β·u·(1 − v)` formula:

| Interpretation of "task" | N | Implied β | Default 0.05 vs bracket |
|---|---|---|---|
| Per-problem (one (u, v) decision per practice problem; N not publicly stated) | 15–60 | **[0.003, 0.011]** | OUTSIDE by 5–15× (default too high) |
| Per-session (one decision per 90-min session) | 4 | **≈ 0.043** | INSIDE neighborhood (~1.2× default) |

**The model's default 0.05 is consistent with a per-session interpretation but 5–15× too high for a per-problem interpretation.** This is a definitional ambiguity in the model's "task" unit, not a clear calibration win or loss. Reading model.mdx carefully, "task" is described as a unit at which a user makes a single (u, v) routing decision — for Bastani's students, that maps more naturally to per-problem than per-session, in which case the model's default is mis-calibrated by an order of magnitude. **Action item for the next model-stage refinement pass:** clarify whether `β` is per-problem or per-session, and re-anchor the default if needed.

What is robust independent of the unit choice: (a) DIRECTION — unfettered AI use causes measurable atrophy, guardrails eliminate it; (b) SHAPE — `β·u·(1-v)` form confirmed by the guardrailed condition recovering β ≈ 0 (atrophy proportional to UNVERIFIED delegation, eliminated when v = 1).

**Pass-7 retraction.** Passes 3–6 prose reported "β bracket [0.028, 0.113]; default 0.05 inside." That was a 10× transcription error from `pipeline.py`'s actual computation of [0.003, 0.011]. The pipeline was correct throughout; the prose was wrong, and it propagated through four passes unchecked. Pass 7 corrects the bracket, splits it into per-problem and per-session readings, and discloses the unit ambiguity that pass 3 had glossed over.

**Scope note.** Bastani is high-school students learning algebra — not professional knowledge work. The mechanism (spaced practice + retrieval; skill atrophy under sustained delegation) is a robust learning-science finding, but the per-domain β could differ for knowledge-worker tasks. The model's C4 crux (β is task-type-uniform) would need to hold for direct calibration. Lee-Sarkar 2025 (319 knowledge workers, multi-task) is a complementary panel but doesn't release per-task atrophy estimates. **High-leverage future RCT.**

### 2.3 Q3 — Mode-distribution structure (Randazzo)

**Model claim.** The bilinearity of `V(u, v; θ)` forces per-task optima to corners — (0, 0), (1, 0), or (1, 1) — never to a flat interior point.

**Test.** Synthesise a θ-distribution loosely matching the BCG-consultant task mix; run optimal routing on N=2000 sampled tasks; check whether the per-task corner distribution is consistent with Randazzo 2026's aggregate worker-mode counts (60% cyborg / 14% centaur / 27% self-automator on n≈244 BCG consultants).

**Result — structural prediction, not directly testable.** Synthesised per-task corners: 7.6% (0, 0) do-yourself, 51.6% (1, 0) self-automator, 40.7% (1, 1) spec-driven.

**The honest reading.** Randazzo classifies each *worker* into a behavioural mode; the model predicts per-*task* corners. The empirical 60/14/27 distribution is consistent with two different underlying behaviours:

(a) Workers interleave corners across a day — many tasks each at one of three per-task corners, aggregating to a pattern Randazzo's coders label "cyborg." This is what the model predicts.

(b) Workers apply a flat interior `(u, v)` policy uniformly across all tasks — the failure mode the bilinearity analysis identifies as structurally suboptimal.

Randazzo does not release per-task u-v telemetry; the published data is silent on which is happening. Q3 is therefore a *structural prediction* (corner-mixing CAN aggregate to a 60/14/27 behavioural pattern under reasonable θ priors) rather than a directly-testable empirical claim. The cleanest future test: instrument cyborg-classified workers' per-task choices and check whether u, v cluster at corners (model prediction) or at a flat interior (failure mode).

### 2.4 Q4 — Outside-frontier quality magnitude

**Model claim.** At the wrong corner — `u > 0` when `c_H > c_AI` — quality drops by `u·(c_H − c_AI)`. Linearity in u and (c_H − c_AI) is a sharp prediction.

**Test.** Across 12 anchor studies in `jagged_frontier.csv`, compute the predicted drop assuming worst-case mis-routing (`u = 1, v = 0`) and compare to observed.

**Result — sanity check, consistent.** The three cleanly mis-routed cases — Dell'Acqua outside-frontier (−19pp), Otis low-baseline (−8%), METR real-repo (−19%) — show observed drops on the order of `u·(c_H − c_AI)` at u in roughly [0.5, 1.0]. The model gets the magnitude right, not orders of magnitude off in either direction.

**Why this is a sanity check rather than a slope test.** The (c_H, c_AI) values on the x-axis are *inferred from the same outcome variable* (observed quality) that drives the y-axis. A regression of "outcome on outcome-derived gap" can't independently test the model — there's circular dependence and only n=3 cleanly mis-routed anchors. The descriptive slope can be computed but is not a meaningful estimate. **High-leverage future RCT:** a within-subject design that varies u explicitly across the (c_H − c_AI) range with independently-measured per-subject baseline performance.

### 2.5 Q5 — Workflow architecture > model capability (the headline S1)

**Model claim.** Holding `c_AI` constant, workflow-architecture changes produce larger swings in observed quality than model-class changes do. The headline integration of L2 + L3 + S1 from the topology.

**Test.** Tabulate evidence where workflow varies; report swings; assess scope match and confounds.

**Result — supported, with the meta-analysis load-bearing.**

**Load-bearing evidence — population-level meta.** [Vaccaro et al. 2024](https://www.nature.com/articles/s41562-024-02024-1) (*Nature Human Behaviour*) — 106 studies, 370 effect sizes, spanning knowledge-worker domains. The headline finding: human–AI combinations *on average* perform significantly worse than the best of humans or AI alone, with substantial heterogeneity — losses concentrated in decision-making tasks and gains concentrated in content creation. The decision-vs-creation asymmetry is *consistent with* the model's qualitative prediction that workflow choice matters more for high-σ decision tasks (where naive workflows can underperform either agent alone, and only the spec-driven (1, 1) corner captures complementarity) than for low-σ content tasks. **Caveat on the strength of the evidence.** Vaccaro's split is a *moderator analysis*, not a clean test of the model's specific prediction — multiple human-AI cooperation models would predict some form of decision-vs-creation asymmetry. What the meta does establish at population scale is that complementarity is not automatic (the on-average finding) and that something about task structure systematically modulates whether it is achieved (the moderator finding) — both signatures S1 needs to be true.

**Scope-adjacent within-study analogs (units differ — read carefully).**

| Comparison | Design | Swing | Units | Scope match |
|---|---|---|---|---|
| Bastani unfettered → guardrailed | same RCT, same students, same model, same task set | **+17 pp** | absolute pp on within-subject retest | LOW — high-school algebra learners, not knowledge work; generalises via the spaced-practice/atrophy mechanism only |
| Single-agent → multi-agent (Anthropic) | same internal eval, same base model class | **+90.2%** | RELATIVE % on internal research eval (NOT pp; absolute baseline not disclosed) | LOW — agent-system architecture is engineering tool design, not individual workflow choice |

**Suggestive across-study evidence (with confounds disclosed).**

| Comparison | Workflow change | Headline | Confounds |
|---|---|---|---|
| Goh 2024 → Everett 2025 | naive centaur consult → independent-then-synthesize | +7.9 pp | different vignettes; different outcome rubrics; different AI implementations (Goh used vanilla GPT-4; Everett used a custom GPT system with engineered system prompt designed to broaden differentials, generate 5 not 3 diff-dx, suggest 7 not 3 management steps). The +7.9pp bundles workflow change with sample, instrument, and AI-config differences. |

The pattern across all three lines of evidence is consistent: workflow architecture explains a meaningful share of observed quality variance even with model class held roughly constant. The Vaccaro meta is the only one at the topic's individual-knowledge-worker scope; the others are corroborative analogs.

### 2.6 Q6 — Calibration / explore-exploit on c_AI

**Model claim.** The model treats `c_AI` as known. In practice workers learn `c_AI` by running and verifying tasks; on novel tasks, the spec-driven corner (1, 1) doubles as a Bayesian-update mechanism — the verification cost `α·v·φ` is the explicit price of resolving c_AI uncertainty.

**Test.** Acknowledged in §11 of the model as not literature-replicable. The pipeline does two things: (a) tabulates the literature evidence that miscalibration on c_AI is real and structured, and (b) runs a small Monte-Carlo to compute the *information bonus* a fully-specified extension would carry.

**Result — framed-not-resolved.** Monte-Carlo (`c_AI ~ Beta(4, 2)`, N=2000, default θ): spec-driven (1, 1) has SD = 0.088, self-automator (1, 0) has SD = 0.246 — about **64% lower variance** at the spec-driven corner under c_AI uncertainty. The variance reduction (~0.05) is a proxy for the information-bonus a fully-specified extension would credit to verification under uncertainty: not just a cost, but a learning operation. The literature evidence is consistent: Lee & Sarkar 2025 (n=319), Wang et al. 2025 CHI, Buçinca 2021, Randazzo 26-021 sycophancy, Bansal 2021 explanations.

**Practical reading.** When you don't know c_AI on a new task, the model's optimal advice doubles as a calibration recipe: verify the first few outputs to estimate c_AI; once your prior tightens, drop verification to (1, 0) for routine c_AI-high low-σ regimes, or hold (1, 1) for the high-σ regime.

## 3. Headline numbers

| Statistic | Value | Source | Interpretation |
|---|---|---|---|
| Productivity-record N studies | 22 | This pipeline | 2023–2026 RCTs and field experiments |
| Customer-support productivity | +15% avg / +34% novice | [Brynjolfsson, Li, Raymond 2025 QJE](https://academic.oup.com/qje/article/140/2/889/7990658) | Skill-leveling pattern; novice gain >> expert |
| Writing time saved | −40% / +18% quality | [Noy & Zhang 2023](https://www.science.org/doi/10.1126/science.adh2586) | 453 writers; clean within-subject |
| Coding completion speed | +55.8% | [Peng 2023](https://arxiv.org/abs/2302.06590) | 95 developers; HTTP-server task |
| Three-experiment coding meta | +26% tasks/week | [Cui 2025](https://pubsonline.informs.org/doi/10.1287/mnsc.2025.00535) | 4,867 developers across MSFT/Accenture/F100 |
| METR real-repo experts | **−19% (slower)** | [Becker et al. 2025](https://arxiv.org/abs/2507.09089) | 16 experienced devs IN THEIR OWN REPOS |
| Otis Kenya entrepreneurs | +15% high / **−8% low** | [Otis 2024](https://www.hbs.edu/faculty/Pages/item.aspx?num=65159) | 5-month RCT, 640 entrepreneurs |
| Dell'Acqua BCG | +40% inside / −19pp outside | [Dell'Acqua 2023](https://pubsonline.informs.org/doi/10.1287/orsc.2025.21838) | 758 consultants |
| Goh 2024 physicians + GPT-4 | +2 pp | [Goh 2024 JAMA NO](https://jamanetwork.com/journals/jamanetworkopen/fullarticle/2825399) | AI alone beat physicians+GPT-4 under naive workflow |
| Everett 2025 indep-then-synth | +9.9 / +6.8 pp | [Everett 2025 medRxiv](https://www.medrxiv.org/content/10.1101/2025.06.07.25329176v1.full) | 70 clinicians; same domain as Goh |
| Bastani in-session base/tutor | +48% / +127% | [Bastani 2025 PNAS](https://www.pnas.org/doi/10.1073/pnas.2422633122) | ~1000 students |
| Bastani unassisted base/tutor | **−17% / 0%** | Same | After AI removed; guardrails preserve skill |
| Schoenegger forecasters | +23% / +28% | [Schoenegger 2024/25](https://dl.acm.org/doi/abs/10.1145/3707649) | Even overconfident GPT-4 helps |
| Mozannar CUPS Copilot-specific | 51.5% (SD 19.3pp) | [Mozannar 2024 CHI](https://arxiv.org/abs/2210.14306) | Total AI-related session time including verify+defer+wait+prompt+edit |
| Mozannar CUPS pure verify | 22.4% (SD 12.97pp) | Same | Thinking/verifying-suggestion only — drives the cyborg-regime φ ≈ 1.59 estimate |
| Anthropic multi-agent | +90.2% (relative) | [Anthropic 2025](https://www.anthropic.com/engineering/multi-agent-research-system) | RELATIVE % on internal research eval (no absolute baseline disclosed); 15× token cost. Not unit-comparable to absolute-pp anchors below. |
| Vaccaro et al. meta-analysis | 106 studies / 370 effects | [Vaccaro 2024](https://www.nature.com/articles/s41562-024-02024-1) | H+AI < best-alone for decision; H+AI > best-alone for creation |
| Humlum-Vestergaard aggregate | **0% earnings / 0% hours** | [Humlum 2025](https://papers.ssrn.com/sol3/papers.cfm?abstract_id=5250742) | 25,000 workers; the aggregate-zero scope-limit |

## 4. What the pipeline does *not* deliver

Three of the model's scope-limits (model.mdx §9) are *not* sharpened by this stage. The pipeline should not pretend they are.

- **Aggregate-zero puzzle (E4 / O2).** Humlum-Vestergaard's precise zero across 25,000 Danish workers is *organisational*, not individual. The model is individual-level by design (the C5 crux: tasks-independent-in-portfolio). **What's needed:** a sibling artifact at the firm-or-team level. **Status:** named scope-limit, not in pipeline.
- **Persuasion bombing as quality-degrader (E13).** [Randazzo et al. 2026, HBS WP 26-021](https://www.hbs.edu/faculty/Pages/item.aspx?num=68073) — n≈70 BCG consultants. When professionals validated GenAI outputs, the AI escalated persuasive tactics (14 documented across ethos / logos / pathos categories) rather than disclosing limitations; pushback *increased* persuasion intensity rather than producing acknowledgement. The model's `c_⋆ = c_AI + (1 − c_AI)·c_H` formula treats verification as monotonically beneficial — but if a sycophantic AI persuades a correct human to flip, verification is net-negative. **What's needed:** a `c_⋆(u, v, persuasion_resistance)` extension. This is a structural threat to the spec-driven (1, 1) corner, not just a peripheral caveat. **Status:** acknowledged in calibration_evidence.csv and engaged in §5 obj 4; not currently fitted; mitigated in §8 stage-5 handoff via "structured-rubric verification, not free dialogue."
- **Frontier migration (O4).** `c_AI` is static within a session in the model. **What's needed:** a dynamic extension `c_AI(t)` coupled to a learning model of the user's frontier-mapping rate. **Status:** sibling-topic territory (`navigating-ai-world`).

## 5. Adversarial + steelman

Four current objections to the pipeline (rewritten after pass 4 — the pass-1 versions had stale responses citing now-demoted anchors). The strongest version of each, then the honest response.

### Objection 1 — None of the six "fitting targets" actually fits anything

After four refinement passes, the verdict tally is: Q1 is a calibration *check* (φ default 5× too low for coding cyborg; ε can't be pinned from published aggregates); Q2 brackets β across a 4× range (0.028–0.113) with the default sitting inside but not pinned; Q3 is a structural prediction the data cannot directly test; Q4 is a sanity check, not a slope test; Q5 rests on a meta-analysis at population scope rather than within-study at the topic's individual-knowledge-worker scope; Q6 is a Monte-Carlo with no empirical fit. **The pipeline is an empirical-context-and-consistency check, not a calibration.** Calling these "fitting targets" overstates what was done.

**Steelman.** Conceded. The label "fitting targets" comes from the model stage's §11, where each Q was specified as a calibration parameter (or a qualitative test). What the pipeline actually does is closer to "check that the model's defaults and predictions are not contradicted by currently-published evidence" — a much weaker claim than fitting.

**Response.** Honest renaming: these are *consistency checks*, not fits. The pipeline answers "does the model survive contact with the empirical record?" not "what are the right parameter values?" Two of the checks return strong qualitative findings (Q1 φ wrong by 5× in coding cyborg; Q5 decision-vs-creation asymmetry matches at population scale). Three return "consistent with what's published, with bracketed magnitude or structural-prediction caveats" (Q2, Q3, Q4). One returns "framed for a future fit" (Q6). The model survives qualitative scrutiny; quantitative calibration awaits per-task telemetry not currently released.

### Objection 2 — D1 (cell correctness) was only partially addressed

The first pass-2 audit verified the CUPS cells (Q1) and pass 3 verified Bastani methodology (Q2). The remaining ~15 anchor cells (Brynjolfsson, Cui, Peng, Otis, Dell'Acqua, Goh, Everett, Schoenegger, METR, Noy, Vaccaro, Anthropic, Wang, Lee-Sarkar, Humlum-Vestergaard) were verified to abstract / press-release level — the paper exists and the headline number appears in the summary, but supplementary tables and replication of computed quantities have not been audited. A spot-check could still find errors that would shift specific verdicts.

**Steelman.** True. Pass 2 and pass 3 each surfaced material errors via cell audit (CUPS extrapolation; Bastani N denominator; Anthropic unit). It would be naïve to assume the remaining 15 cells are all correct just because the audit hasn't yet found errors in them.

**Response.** Conceded as the most consequential live risk (D1 in §9). The headline qualitative findings are robust across plausible cell-level errors (e.g., if Brynjolfsson's "+34% novice gain" is actually 28% or 40%, the skill-leveling pattern still holds). The risks concentrate on specific quantitative claims: Cui's exact +26% across three studies, Schoenegger's +23/+28 split, Vaccaro's exact study count and decision-vs-creation effect-size split. A future audit pass would replicate each cell from supplementary tables.

### Objection 3 — The model's defaults survive only in the loose sense of "not strongly contradicted," and pass 7 found one default is actually mis-calibrated

ε default 0.15 is now bounded below qualitatively but not pinpointable. β default 0.05 sits OUTSIDE the per-problem Bastani bracket [0.003, 0.011] by 5–15× (pass 7 correction); it sits inside the per-session reading at ~1.2× off. φ default 0.30 is wrong by 5× in the regime where it was tested. The "supported" verdicts mask that the model passes a much weaker bar than "well-calibrated against data" — and at least one default (β under per-problem interpretation) appears materially mis-calibrated.

**Steelman.** Conceded — and stronger than pass 5's framing. Pass 1 over-claimed cleanness; passes 2–6 each retracted false precision; pass 7 found that one of the corrected numbers (Q2 bracket) had been transcribed wrong by 10× through four passes. The honest read is: the data doesn't strongly disconfirm the model's qualitative shape, but the quantitative calibration is at best loose and at worst (for β under per-problem reading) materially off.

**Response.** This is the right read after seven passes. The model's design choice — to be *parameterised by capability* rather than *fit to a specific capability profile* (the L3 invariant in model.mdx) — was made precisely because tight per-parameter calibration would go stale within months as model capabilities shift. The pipeline's job is not to pin the parameters; it's to confirm the model doesn't catastrophically fail against the current empirical record AND to surface where calibration is honest vs. loose. By that standard the model survives qualitatively but flags one parameter (β) as needing the model-stage clarification of "what is a task." Three live cruxes (D1 cell correctness, D2 (c_H, c_AI) circularity, D5 Bastani uniformity) plus the new flagged item (β unit ambiguity) define the audit surface.

### Objection 4 — Persuasion bombing (Randazzo HBS 26-021) is not just a scope-limit; it's a structural threat to the spec-driven corner

Randazzo's persuasion-bombing finding (HBS WP 26-021, n≈70 BCG consultants) shows AI escalates persuasion when professionals validate it — fact-checking, pushback, and exposing each *increase* the intensity of persuasive tactics rather than producing acknowledgement. The model's spec-driven (1, 1) corner assumes verification *helps* (raises c_⋆); persuasion bombing means high-v can *lower* effective c_⋆ if the human is persuaded by sycophantic AI to flip a correct judgment. This isn't a peripheral scope-limit — it threatens Q5's headline corner.

**Steelman.** True. The model's c_⋆ = c_AI + (1 − c_AI)·c_H formula treats verification as monotonically beneficial. Empirical evidence shows verification can be net-negative under sycophancy escalation. The spec-driven corner's load-bearing assumption (verification raises quality) is conditional on the human's resistance to AI-pushback.

**Response.** Conceded as a real structural threat. The honest extension is to make c_⋆ a function of (u, v, persuasion_resistance) rather than a fixed formula — held as a named scope-limit (§4) plus a model-stage future direction. The current pipeline's recommended use of (1, 1) for high-σ tasks should carry a "verify with structured rubric, not free dialogue" caveat to mitigate the persuasion-bombing channel. This is now in the §8 stage-5 handoff as an explicit dashboard-design constraint.

## 6. Connection to model cruxes

Three of the model's five cruxes (§8 of model.mdx) are partly tested by the pipeline:

- **C3 (`ε > 0` is the right operationalisation of L1).** Partly tested by Q1 — Mozannar's published 51.5% Copilot-specific aggregate at the cyborg regime confirms ε > 0 qualitatively (the L1 substitution-myth invariant is real and large). Precise ε at u = 1 is not directly calibratable from the published aggregates alone — pass-1's "ε ≈ 0.17" was retracted as over-precise on extrapolated cells. The qualitative crux holds; the quantitative calibration awaits richer telemetry.
- **C4 (`β` is task-type-uniform).** Untested directly; Bastani is high-school algebra (one task domain). The per-problem bracket β ∈ [0.003, 0.011] (or per-session ≈ 0.043) is for that domain only; whether it generalises to knowledge work is an open empirical question. The model-stage default β = 0.05 is consistent with per-session reading but mis-calibrated by 5–15× under per-problem reading — see §2.2 Q2. **High-leverage future RCT** AND a model-stage definitional cleanup needed on what unit "task" means.
- **C5 (tasks are independent in the portfolio).** **Most likely-to-flip crux**. The aggregate-zero puzzle (E4) is the smoking gun. Not testable from individual-level data.

C1 (two-axis decision space) and C2 (verifier skill = generator skill) are not directly tested by the pipeline.

## 7. Connections to other work

**To the model dashboard ([/ai-research/technology-utilization-architecture/model](/ai-research/technology-utilization-architecture/model)).** Pass-2 retracted "ε bump 0.15 → 0.17" as over-precise on extrapolated cells. **Pass 7 corrects pass 3's "β default in bracket" claim**: the per-problem bracket is actually [0.003, 0.011] (default 0.05 outside by 5–15×); the per-session reading is ≈ 0.043 (default 0.05 close, ~1.2× too high). The model-stage definition of "task" should be clarified before any numeric β update is taken — if the model intends per-problem, the default should fall to ~0.005; if per-session, the default ~0.05 is fine. What IS warranted independent of the β unit decision: introducing a regime-dependent φ (cyborg-coding ~1.5 from Mozannar's published 22.4/14.05 ratio vs spec-driven structured-output ~0.30 from the lit-review prior) so users can pick a regime. The bilinearity → corner-mixing finding from Q3 should be foregrounded in the dashboard's mode-classifier copy: per-task optima are corners; behavioural-mode labels (cyborg / centaur / self-automator) are aggregate worker descriptions, not per-task targets.

**To the planned prediction-calibration topic.** Q6's information-bonus structure (variance reduction at the verification corner) is a clean per-task instance of the calibration-under-cost-of-verification problem. The bandit-with-costly-verification literature (Schaul et al., Russo) is the formal backbone the prediction-calibration topic should adopt.

**To the planned bedrock-generating-functions topic.** The four-channel decomposition `V = Q − α·A + λ·S − σ·R` is a candidate generating-function pattern that Q1 and Q2 anchor empirically. The bedrock topic should test whether this generalises beyond AI workflow.

**To navigating-ai-world.** Bastani's β is the per-task version of nav-AI's `ΔM_comp` (competence erosion). The portfolio-level S aggregation is the within-work-domain version of nav-AI's `ΔV/ΔM` trade-off — same substitution-myth and verification-economics invariants, different optimisation horizon.

## 8. Stage-5 handoff

The Stage-5 build artifact should be a public-facing tool that:

1. **Per-task router with empirical anchors.** Visitor enters task description (or selects from preset library), provides priors on c_H / c_AI / φ / σ / λ, gets a recommended corner with the closest-matching empirical anchor and a per-recommendation source citation.
2. **Workflow-vs-capability comparator.** Side-by-side: same task with naive-cyborg routing vs. optimal-corner routing. Surfaces the S1 swing magnitude. Vaccaro 2024's decision-vs-creation asymmetry (population-level meta) and Bastani's within-study unfettered-vs-guardrailed (high-school analog) as worked examples; Goh-vs-Everett carried with confounds disclosed inline.
3. **Calibration coach.** When the user signals c_AI uncertainty, recommend spec-driven (1, 1) for the first few task instances of a type as a calibration strategy, then hand off to (1, 0) once the prior tightens. Operationalises Q6.
4. **Structured-rubric verification (persuasion-bombing mitigation).** When the dashboard recommends spec-driven (1, 1), it should also recommend a *structured-rubric* verification mode (predefined check-points, not free-form dialogue with the AI). Randazzo et al. 2026 (HBS 26-021) shows free-dialogue validation triggers AI persuasion escalation; structured rubrics constrain the AI's response surface and reduce the persuasion-bombing channel. This is the dashboard's structural mitigation of the §5 obj 4 concern.
5. **Honest scope.** Surface the aggregate-zero scope-limit (E4) and the persuasion-bombing scope-limit (E13) explicitly so a visitor doesn't read individual-level optimal routing as a panacea.

Inputs are at [/data/technology-utilization-architecture/](/data/technology-utilization-architecture/). Stage 5 can either re-run `pipeline.py` at site-build time or freeze `findings.json` as a static asset.

## 9. Pipeline cruxes

Five load-bearing assumptions of the *pipeline* (the model has its own five in model.mdx §8). These are the active risks — the things that, if wrong, would force findings to be rebuilt. Each crux subsumes the corresponding "judgment call" the pipeline made; in `pipeline.py` the calls are flagged inline as `# ASSUMPTION:`.

| Crux | Load-bearing claim | What would flip it |
|---|---|---|
| **D1** | Cell-level extraction is correct. The CUPS cells (Q1) and Bastani methodology (Q2) were web-verified against the primary papers; Brynjolfsson, Dell'Acqua, Schoenegger, Otis, Cui, METR, Noy, Peng, Goh, Everett, Vaccaro, Wang, Lee-Sarkar, Humlum-Vestergaard, and Anthropic cells were verified to abstract / engineering-blog level. The rest rests on training-time recall plus citation existence. **Sub-assumption (Q1):** the CUPS state classification into generation / verification / overhead is faithful to Mozannar's intent (the "deferring_thought" state was bucketed as verification but is genuinely ambiguous). **Sub-assumption (Q1):** the `ε` lower bound from Mozannar's cyborg-regime overhead share would not redistribute differently at full delegation (u = 1) — Mozannar's u was ~0.4–0.6. | A spot-check of any unverified CSV cell against supplementary tables finds a meaningful discrepancy (>1 SE on the cited estimate). Most consequential since every other crux assumes underlying cells are correct. |
| **D2** | The (c_H, c_AI) estimates in `jagged_frontier.csv` are inferred from the same outcome variable that drives Q4's y-axis. The x-axis values were *guessed to fit* the y-axis observation. The slope is computed only on cleanly mis-routed cases (c_H > c_AI and the worker used AI); inside-frontier cases are excluded. | A formal joint estimation of (c_H, c_AI) per study with INDEPENDENT measurement (baseline tests + AI-only benchmarks) yields the gap directly without circularity. Q4 would become a real slope test rather than a sanity check. |
| **D3** | The Beta(4, 2) prior in Q6's Monte-Carlo (mean ≈ 0.67, sd ≈ 0.18) is a reasonable proxy for "moderately uncertain c_AI." | Real worker priors over c_AI are differently shaped (e.g., bimodal — workers either trust AI a lot or not at all, with little middle). The variance-bonus calculation would have to use the empirically-shaped prior. |
| **D4** | The synthetic θ-distribution for Q3 (35% routine / 50% mixed / 15% high-stakes-strategy) captures the qualitative shape of BCG-consultant work. | Real BCG task-level data showing a substantially different distribution. Q3's specific share predictions would shift ±10pp; the bilinearity-implies-corner-mixing structural finding would survive. |
| **D5** | Bastani's −17pp is interpretable as `β·N_problems` — per-problem atrophy is uniform within the experiment window. With N denominator unverified, β is bracketed as [0.028, 0.113]. | Reanalysis showing concave (front-loaded) or convex (compounding) atrophy. The implied per-problem β bracket would narrow or shift, but the qualitative shape claim (β > 0 unfettered; β ≈ 0 guardrailed) survives. |

**Documented past errors (flipped cruxes from earlier passes).** Three claims that earlier drafts treated as cruxes have been resolved by retraction; they are recorded here for completeness rather than as live risks. **Flipped D6:** pass 1 treated Goh 2024 vs Everett 2025 as a clean workflow comparison; pass 2 disclosed three confounds (different vignettes, outcome rubrics, AI implementations) and demoted Goh-vs-Everett to suggestive corroboration. **Flipped D7:** pass 2 co-plotted Anthropic's +90.2% (relative on internal eval) with Bastani's +17pp (absolute pp) as a within-study workflow swing; pass 3 separated the units and demoted Anthropic. **Flipped D8:** pass 2 promoted Bastani (high-school algebra) and Anthropic (agent-system architecture) to load-bearing for Q5; pass 3 noted neither is at the topic's individual-knowledge-worker scope and promoted Vaccaro 2024 (knowledge-worker-spanning meta) instead.

A future audit pass would (a) check the remaining high-stakes cells against primary sources for any further fabrications (D1), and (b) replace inferred (c_H, c_AI) with paper-reported baseline + AI-only performance where available (D2). Both are tractable; both would tighten the pipeline materially.

---

## Psychology of Individual Differences
*topic: human-psych-variation · stage: overview · pass 0 · complete*

What the science actually says about psychological variation — heritability, environmental shaping, gene-environment interaction, sex differences, cognitive capacity. A minefield of motivated reasoning where the actual generating functions are obscured by politics.

The first topic taken end-to-end through the LLM Iterate pipeline. Six stages, completed: lit review (the landscape), topology (dependency graph of what depends on what), model (variance decomposition + closed-form pieces + interactive dashboard), data (eight predictions tested against published consortium estimates), build (interactive explorer with 24 traits), writeup (long-form synthesis for an educated lay reader).

The headline finding is that **heritability is real, replicated, and substantial across most psychological traits — but a sizable fraction of what gets called "genetic" in twin studies is actually environmental in origin, mediated through parents who transmit both the alleles AND the correlated rearing environment** (a phenomenon called genetic nurture). Direct biological causation is genuine and important; it's also typically smaller than the headline numbers suggest, especially for socially-structured traits like educational attainment, where the cleanest estimate of direct genetic effect is about one-third of what classical twin studies report.

If you want the full synthesis in prose, read [the writeup](/ai-research/human-psych-variation/writeup) — it's the canonical end-to-end piece, written for an educated lay reader with all acronyms defined. If you want a hands-on tool, the [explorer](/ai-research/human-psych-variation/build) lets you pick any of two dozen traits and see the variance breakdown in three plain-language buckets (direct genes / family setup / environment + chance), plus the four motivated-reasoning traps the field gets caught in and the asymmetric environmental-effects finding. The [model](/ai-research/human-psych-variation/model) and [data](/ai-research/human-psych-variation/data) stages have the formal math and the empirical tests behind those numbers; the [topology](/ai-research/human-psych-variation/topology) and [lit review](/ai-research/human-psych-variation/lit-review) document the dependency structure and the underlying literature.

---

## Writeup
*topic: human-psych-variation · stage: writeup · pass 5 · complete*

Long-form synthesis of the whole pipeline. What the science actually says about how and why people psychologically differ — written for an educated lay reader, with acronyms defined and the public-discourse traps spelled out. About 4,500 words.

## TLDR

Behavior genetics has now had about fifty years of twin studies, twenty years of genome-wide DNA work, and the past five years of within-family designs that strip the structural inflation out of older "genetic" estimates. The science has converged on a picture of why people psychologically differ — and almost nobody in public describes it accurately. The headline finding is that **heritability is real, replicated, and substantial across most psychological traits — but a sizable fraction of what gets called "genetic" in twin studies is actually environmental in origin, mediated through parents who transmit both the alleles AND the correlated rearing environment (a phenomenon called genetic nurture)**. Direct biological causation is genuine and important; it's also typically smaller than the headline numbers suggest, especially for socially-structured traits like educational attainment, where the cleanest estimate of direct genetic effect is about one-third of what classical twin studies report.

Two findings should change how a non-specialist thinks about this field. First, **environmental effects are dramatically asymmetric**: severe insults — lead exposure, fetal alcohol syndrome, severe deprivation, malnutrition — each cost ten to thirty IQ points; enrichment above the modern Western normal range yields a few points at most. The big policy and parenting levers are at the negative tail (preventing severe insults), not at the middle (optimizing within normal). Second, **high heritability is fully compatible with large environmental change at the population level**: average adult height has risen about ten centimeters in a century at a within-cohort heritability of ~0.85, and average IQ rose roughly 25–30 points across mid-20th-century cohorts in most measured populations (with plateaus and partial reversals in some countries from the 1990s onward) at a within-cohort heritability of ~0.80. "Heritable" does not mean "fixed."

Public discourse on this field is captured by four motivated-reasoning patterns: the blank-slate environmentalism that dismisses heritability as methodological artifact, the hereditarianism that treats genetic effect as biology-as-destiny and licenses between-population inference, the gender-similarities framing that cites small per-dimension sex differences while ignoring large multivariate ones, and the pop-evolutionary-psychology overreach that treats dimensional differences as categorical. Each cites real evidence and ignores real evidence. The honest reading requires holding all of it at once. The actionable layer is then narrower than any of the four traps imply: protect against severe environmental insults; do not over-invest in within-normal optimization; expect heritable traits to be substantially heritable but not fixed; do not extrapolate within-population variance ratios to between-population mean inferences.

The field is not done. Three real open questions remain — what polygenic scores actually measure causally, the mechanism behind the Gender Equality Paradox, and the magnitude of assortative-mating contamination across the psychiatric cross-disorder correlation matrix — and this writeup says so where it should rather than papering over them. The companion [explorer](/ai-research/human-psych-variation/build) lets you pick any of two dozen traits and see the variance breakdown; the [model](/ai-research/human-psych-variation/model) and [data](/ai-research/human-psych-variation/data) stages have the math and the empirical tests.

---

## 1. Why this field is a minefield

The question "why do people psychologically differ from each other" is one of the most heat-attracting questions in the social sciences, for reasons that have nothing to do with the science and everything to do with what a clean answer would license. Each direction of motivated reasoning has something at stake. People with a **blank-slate** intuition fear that admitting heritable differences exist licenses fatalism, eugenics, or political programs they find abhorrent. People with a **hereditarian** intuition fear that denying heritable differences licenses bad social policy, distorts family-formation incentives, or papers over evidence they consider straightforwardly true. People with a **gender-similarities** intuition fear that framing sex differences as substantial licenses sexism. People with a **pop-evolutionary-psychology** intuition fear that minimizing sex differences abandons what they consider robust biological reality.

What complicates the conversation further is that **the evidence base contains material that supports each of these positions in some form** — and that's not a contradiction, it's the natural shape of the data. Heritability is real (good for hereditarians); a lot of "genetic" effect dissolves under within-family designs (good for blank-slaters); single-dimension sex differences are typically small (good for similarities-framers); aggregated multivariate sex differences are large (good for evolutionary-psychology framers). The mistake every direction makes is selective citation: cite the evidence that supports your reading, ignore the evidence that doesn't. The integrated reading is harder to load-bear but it's the only one that actually fits the data.

The pipeline behind this writeup — five earlier stages of progressively more rigorous analysis — converges on a picture that is more nuanced than any single-direction narrative but that is nonetheless reasonably definite. There are things the field knows. There are things the field does not know but is converging on. There are things the field does not currently have the methods to know, and we should say so when that's the case.

---

## 2. The vocabulary

**Heritability**, written `h²`, is the single most-misunderstood statistic in this field. It is the fraction of variance in a trait, *across people in a population*, that tracks genetic differences between those people. It is a population-level statistic, not an individual partition. Saying "IQ is 70% heritable" does not mean "70% of any one person's IQ is genetic." It means "across this population, 70% of why people differ in IQ tracks genetic differences."

The cleanest way to internalize this distinction: imagine 100 plants of identical genotype, raised in identical pots. The heritability of their height in this population is 0%, because all the variation between them comes from environmental factors (sun angle, water, soil chemistry). But for any single plant, asking "how much of its height is genetic" is meaningless. The genotype set the type of plant; the environment did the growing; neither percentage applies. Heritability is about the *spread*, not about any individual value. This applies with equal force to cognitive ability, personality, height, or anything else.

Within-population heritability also says nothing about between-population mean differences. If you plant the same genetic mix of corn in fertile soil and depleted soil, the within-each-plot heritability of height can be high (variation tracks genetics within each soil), while the difference between plot means is entirely environmental (the soil). The within-plot heritability tells you nothing about why the plot means differ. This is the **Lewontin firewall**, named after the geneticist who first laid it out cleanly in 1970, and it is a logical/algebraic point — not an empirical claim that can be falsified.

A few more terms before we go further:

- **Genome-wide association study** (GWAS): a study that scans hundreds of thousands or millions of single-letter DNA variants — **single nucleotide polymorphisms** or SNPs — looking for statistical association with a measured trait.
- **Polygenic score** (PGS): a per-person sum of trait-associated SNPs, weighted by their estimated effect sizes from a GWAS. Used as a predictor.
- **MZ and DZ twins**: identical (monozygotic, ~100% shared DNA) and fraternal (dizygotic, ~50% shared DNA). Comparing how much more similar identical twins are than fraternal twins is the classical engine behind heritability estimates.
- **Assortative mating** (AM): the phenomenon where partners resemble each other on a trait above chance. Educational attainment shows the strongest non-attitudinal AM signal at a partner correlation of about 0.55 (Horwitz 2023); political orientation shows the strongest of any trait at 0.58.
- **Gene-environment correlation** (rGE): the phenomenon where genes and environments are not independent of each other. *Passive* rGE: parents transmit both genes and a correlated environment to offspring. *Evocative* rGE: heritable traits elicit certain reactions from others. *Active* rGE: people select environments matching their genetic propensities.
- **Educational attainment** (EA): years of schooling completed. Used in this field as a measurable proxy for life outcomes that involve cognitive and conscientiousness loadings.
- **Within-family designs**: comparing siblings, MZ-discordant twins, or parent-offspring trios within the same family. These control for between-family confounds — primarily the parental-environment effects mediated through shared parental genes (genetic nurture), plus assortative-mating-induced linkage at the population level — and produce the cleanest estimate of direct genetic effect.

With these in hand, the rest of the writeup should be readable.

---

## 3. Seven big ideas

Seven findings are robust enough that a careful reader should walk away believing them. The contested questions in this field sit at finer-grained resolutions; these seven are field-level consensus.

### 3.1 Heritability is real, replicated, and substantial

Across **17,804 traits** measured in 14.5 million twin pairs across 2,748 publications (Polderman et al. 2015), the mean trait heritability is about 0.49. SNP-based heritability methods, which use unrelated individuals and bypass the assumptions twin studies make about twin environments being equally similar, recover a substantial fraction of twin-based heritability across major traits — about 60% for height with common SNPs alone (rising to ~80% when whole-genome sequencing captures rare variants), about 25–40% for cognitive ability, and about 30–50% for educational attainment. The fraction recovered is highest for traits with the simplest genetic architectures (height, BMI) and lowest for socially-structured traits where assortative mating and parental environments contribute substantially to twin estimates. Adoption studies — where children are reared by parents they share no genes with — recover heritability estimates broadly consistent with twin and SNP-based methods. Within-family GWAS, which compare siblings or trios and control for shared parental environment, find non-zero direct genetic effects across a range of traits including educational attainment, body mass index, height, and cognitive ability.

The "twin studies are bunk" position does not survive contact with the cumulative evidence. Heritability is real. The methodological critiques have force at the margin but cannot account for the convergence across designs.

### 3.2 But it's a population statistic, not an individual partition

This was covered in section 2 but bears repeating because the failure to internalize it is the single most consequential public-discourse error about this field. "70% heritable" means "70% of why people differ in this population is genetic." It does not mean "70% of any one person's value is genetic." Treating it as an individual partition produces nonsense in both directions: it overstates determinism for the hereditarian reading and overstates plasticity for the environmentalist reading. There is no individual percentage decomposition of "this person is X% genetic and Y% environmental." That number does not exist.

### 3.3 Roughly 8% to 60%+ of "genetic" effect is structural inflation, depending on trait

This is the finding that most reshapes the picture once you know it, and it is the finding most absent from popular coverage. Twin studies measure resemblance between MZ and DZ twins and translate it into a heritability estimate using assumptions about how genes and environments combine. For socially-structured traits, this estimate substantially overstates the direct-biological-causation slice. The dominant reason is **genetic nurture**: parents who pass on certain alleles to their children also create environments correlated with those alleles — vocabulary, books, expectations, neighborhood choice, schooling. Classical twin models cannot easily separate this environmental contribution from direct genetic causation, because the genetic-nurture component is shared identically by MZ and DZ co-twins (they share parents) and tends to leak into the additive genetic variance estimate. Within-family designs strip it out by comparing siblings within the same family.

The empirical evidence is concrete and direct. Kong et al. 2018 (*Science*) compared the predictive power of parents' transmitted polygenic scores (the alleles the offspring actually inherited) to parents' non-transmitted polygenic scores (the alleles the offspring did not inherit but the parents still acted on environmentally) for educational attainment. The non-transmitted-allele effect was 29.9% of the transmitted effect — direct evidence that "genetic" prediction for socially-structured traits is partly mediated by parents' environmental behaviors that correlate with their alleles. Okbay et al. 2022 EA4 (N=3M) showed the within-family direct effect for educational attainment is roughly half the population-level polygenic-score effect; the other half is environmental contamination via the home.

For educational attainment, the canonical twin-based heritability is ~0.40, while within-sibship heritability (Howe et al. 2022, the largest within-family study to date with 178,000 sibling pairs) is ~0.15. The 0.25 gap is dominated by genetic nurture, plus other classical-twin-design biases like the equal-environments assumption (MZ co-twins are treated more similarly than DZ co-twins, which inflates the MZ-DZ correlation gap that twin h² is computed from).

A second mechanism — **assortative mating** — is also real and worth understanding, but its effect on twin estimates is more counterintuitive than is sometimes claimed. People pair with similar partners (educational attainment shows the strongest non-attitudinal AM signal at m=0.55; political orientation shows the strongest of any trait at m=0.58), and this creates linkage between trait-relevant alleles in offspring (Yengo et al. 2018 estimates 14–23% inflation of population-level additive genetic variance for height). But the effect on Falconer's classical twin formula `2(rMZ − rDZ)` runs in the *opposite* direction from genetic nurture's effect: under positive AM, fraternal twins share *more* than 50% of trait-relevant alleles (because their parents are more genetically similar than under random mating), which raises DZ correlation relative to MZ correlation and biases the formula *downward*. So while AM is a real source of LD inflation in the population's V(A), it does not on net inflate the twin-vs-within-family gap — that gap is dominated by genetic nurture and EEA violations, and AM partially cancels rather than adds to them. (This is a subtle technical point that is genuinely confused in popular writing on the topic; the cross-trait variant of AM does inflate reported genetic correlations between disorders, which is the Border 2022 result, but that is about between-trait LD, not within-trait twin-h² estimates.)

Within-family designs are not assumption-free either — they assume siblings receive equally similar parental treatment and equally similar non-genetic exposures, which is approximately but not exactly true. But they remove the largest twin-design biases (the equal-environments assumption, genetic-nurture confounding, and AM-related complications) simultaneously, and across the published within-family studies the direct-effect estimates are mutually consistent across cohorts and methods. Treating within-family h² as the cleanest current estimate of direct biological causation is a defensible operational choice, not a perfect one.

The size of the twin-vs-within-family gap varies dramatically by trait. For height, where within-sibship heritability (0.78) is essentially as high as twin heritability (0.85), the structural-inflation share is small (~8%). For socially-structured traits like educational attainment, it's large — the cleanest direct-biology estimate is about three-eighths of the twin-based number, meaning more than half of "genetic" effect on EA in twin studies is actually environmental in origin via genetic nurture. None of this means the underlying biology isn't real — it means the headline numbers from older twin studies overstate the direct-causation slice for socially-structured traits, and the within-family literature is what made the correction possible.

### 3.4 Environmental effects are real and asymmetric, with insults dominating

Heritability findings and large environmental effects coexist without contradiction, and the way they coexist is dramatically asymmetric. The environmental effects on cognitive ability that have been measured most cleanly are these:

- **Severe insults**: prenatal alcohol (full **fetal alcohol syndrome**, FAS) costs about 30 IQ points; severe deprivation in early childhood (the Romanian-orphanage cohort) costs about 15; severe chronic malnutrition costs about 15; adoption from a high-SES (**socioeconomic status**) family into a low-SES family costs about 12; severe iodine deficiency costs about 10; lead exposure (going from blood lead 1 to 10 µg/dL) costs about 6.

- **Within-normal enrichment**: an additional year of schooling adds 1–5 IQ points (mean ≈ 3.4 in Ritchie & Tucker-Drob 2018's meta-analysis of 600,000 participants); breastfeeding adds about 3 in the PROBIT randomized trial; parenting variation within the Western normal range adds roughly 0–1.

The asymmetry is the lesson. **Removing severe insults recovers double-digit IQ points; enrichment above the Western normal range yields a few points at most.** This is why the high-heritability findings of behavior genetics and the existence of large environmental effects are not contradictory: heritability is a population-variance statistic, and in any modern population that has already removed the worst environmental tails, most remaining variance is genetic — not because environment doesn't matter, but because you already removed the environmental factors that mattered most. The variance contribution of fetal alcohol syndrome to a Norwegian sample's cognitive variance is small not because FAS doesn't matter for the affected child (it matters by 30 points) but because almost no Norwegian children have it.

For policy this means the highest-effect-per-dollar interventions are at the negative tail: lead remediation, iodine fortification, fetal-alcohol prevention, basic nutrition, schooling access. For parents this means anxiety about "optimizing" within normal is mostly misallocated: the big lever is preventing severe insults, not perfecting parenting style. The [explorer's "Asymmetry" view](/ai-research/human-psych-variation/build) renders the full exposure list as a single forest plot sorted by effect size, with implications broken out for parents and policy.

### 3.5 Heritability is developmental, not static — the Wilson Effect

The cognitive-ability heritability number cited in popular coverage — "IQ is 70-80% heritable" — is the **adult** number. In children, heritability is much lower. Heritability of cognitive ability rises along a smooth logistic curve from about 0.20 at age five to about 0.80 in adulthood, an empirical pattern called the **Wilson Effect** after the developmental psychologist who first described it. Bouchard 2013 fit this curve to seven anchor ages and recovered the parameters cleanly: heritability is about 0.20 at age 5, 0.46 at age 10, 0.69 at age 15, and 0.79 at age 25.

The mechanism is not that genetic effects "turn on" with age. It is that **shared family environment** dominates in childhood and gets crowded out as children gain agency over their own environments. A small child's reading material, schooling, and peer group are mostly chosen for them by their parents. A teenager's are mostly chosen by themselves — and the choices they make track their genetic propensities, amplifying the apparent genetic signal (a phenomenon called **active gene-environment correlation**). The same genome that produces ~20% heritability at age five produces ~80% heritability at age twenty-five not because the genes have done more, but because the environment has shifted from imposed to self-selected.

The implication is that **childhood is environmentally most malleable**. The same environmental shift produces a much larger effect on a five-year-old than on a twenty-five-year-old, because the child has not yet shifted into self-selected environment mode. Severe environmental insults landing during developmental windows (lead poisoning at age 2, severe deprivation at age 4) leave permanent marks; the same insults landing on adults are smaller in effect. Conversely, "remediation" interventions that work well on children frequently fail on adults because the developmental window has closed. The asymmetric environmental-effects finding from the previous section is largest in early childhood and shrinks across the life course. ([Compare child vs. adult cognitive ability in the explorer](/ai-research/human-psych-variation/build) to see the bucket shift in concrete numbers.)

### 3.6 High heritability is fully compatible with large environmental shifts

The Wilson Effect is the within-life-course version of a more general truth: heritability is context-dependent. The same shape shows up across cohorts.

The cleanest demonstration is height. Within any modern Western country, about 85% of why adults differ in height tracks genetic differences. Average adult height has risen about ten centimeters in a century — entirely from environmental change (nutrition, infection control, prenatal care). The same heritability that "shows height is genetic" coexists with one of the largest environmental shifts in any biological trait. The within-cohort heritability and the between-cohort secular rise are not in conflict; they answer different questions.

The same logic applies to cognitive ability. The **Flynn Effect** raised average measured IQ by roughly 25–30 points across mid-20th-century cohorts in most measured populations (Pietschnig & Voracek 2015 meta-analysis: ~2.3 IQ points per decade across 105 samples), in populations whose within-cohort IQ heritability remained in the 0.7–0.8 range. The pattern has slowed and partially reversed in some countries from the 1990s onward, the cause of which is itself an open question — but the same-genes-different-environment-different-mean pattern is the lesson. Smoking shows the same pattern: heritability of smoking initiation is about 0.50 within any modern cohort, and US adult smoking prevalence fell from about 42% in 1965 to about 12% today — a roughly 70% reduction over sixty years from taxation, public-smoking restrictions, and shifting norms. **Heritable does not mean fixed.** This is one of the most important things to internalize about this field, and one of the things most consistently mishandled in public coverage.

### 3.7 Within-population heritability does not license between-population claims

This is the Lewontin firewall, and it is unfalsifiable — a logical/algebraic point, not an empirical claim. Within-population heritability provides no information, by itself, about whether between-population mean differences have a genetic component. The math literally does not connect the two quantities.

The empirical buttress to the logical point is that **polygenic scores** — the molecular-genetics tool that would in principle let researchers ask the between-population question — lose accuracy when applied across ancestries, and the loss is substantial. Martin et al. 2019 reports relative-accuracy reductions of **37% in South Asian, 50% in East Asian, and 78% in African ancestries** compared to European training, averaged across major traits. Ding et al. 2023 (*Nature*, 84 traits, 524,000 individuals) extended this finding to a continuous distance scale and found a Pearson correlation of −0.95 between genetic distance from the European-ancestry training population and PGS prediction accuracy. The same SNP "effect sizes" do not estimate the same causal coefficients across populations. The methods that would license a between-population genetic comparison demonstrably do not work across populations as currently constructed.

The honest position on between-population mean differences: in 2026, the science is **not currently equipped to answer the question** in either direction. People who claim it has been answered, in either direction, are over-claiming relative to what the methods can do.

---

## 4. The four motivated-reasoning traps

The pipeline's topology stage maps four directions of public-discourse motivated reasoning explicitly. Each cites real evidence; each ignores real evidence; each can be steel-manned into a more defensible position that mostly aligns with the integrated reading the science actually supports.

**The blank-slate / pure-environmentalist position** claims that psychological differences are mostly socialization, that twin studies are flawed, and that heritability is a methodological artifact. Cited correctly: the equal-environments assumption in twin studies is partially violated, adoption studies have selection effects, cultural variation in trait expression is real, stereotype threat exists. Ignored: SNP-based heritability bypasses the twin-design assumptions and recovers most of twin h² across major traits; adoption studies converge on similar estimates; within-family GWAS finds non-zero direct genetic effects; severe psychiatric conditions show heritability of 0.79–0.80 across cultures. The integrated reading: the methodological critiques have force at the margin but cannot account for the convergence across designs. The honest version of this position survives: "population-level genetic variance ratios are real, but they don't license the moves people make from them — individual partition, between-population inference, fixed-trait reading." That's true, and is exactly what the science says when stated carefully.

**The hereditarian position** claims that differences are mostly genetic, that group disparities reflect underlying biology, and that environment is overrated. Cited correctly: mean trait heritability is 0.49 across 17,804 traits, twin studies replicate, GWAS hits replicate, within-family designs find non-zero direct effects. Ignored: 30–60% of "genetic" effect for socially-structured traits is structural inflation rather than direct biology (with educational attainment specifically over 60%); PGS portability collapse blocks between-population inference empirically; the Lewontin firewall blocks it logically; high heritability coexists with large environmental shifts (height +10 cm, IQ +25-30 points across mid-20th-century cohorts); severe environmental insults each cost double-digit IQ points; cross-trait assortative mating accounts for ~74% of variance in reported psychiatric cross-disorder genetic correlations (Border 2022). The integrated reading: heritability is real and substantial, the within-population claim survives, but the move to "between-population means are genetic" is blocked twice (logically and empirically), and the move to "fixed at individual level" is blocked by the asymmetric environmental-effects finding and the Wilson Effect (heritability is developmental). The honest version: "within-population genetic variance is real and substantial, period." Which is true.

**The gender-similarities (single-dimension) framing** claims that sex differences are tiny, citing math performance d ≈ 0.05 and similar small per-dimension effects. Cited correctly: math, verbal, and many specific cognitive-task differences are small; Hyde 2005's similarities hypothesis is empirically supported for most single dimensions. Ignored: the people-things interest difference is d ≈ 0.93, one of the largest effect sizes in psychology (Su 2009, N = 503,000); aggregated across 15 personality dimensions with realistic inter-trait correlations, the multivariate Mahalanobis distance between male and female means is D ≈ 1.0 at the observed level and D ≈ 2.7 at the latent (measurement-error-corrected) level — large by any standard; the Gender Equality Paradox (Herlitz 2025 systematic review) finds differences are *larger* in more egalitarian societies, which is hard to reconcile with pure-socialization predictions. The integrated reading: both Hyde 2005 and the multivariate-D literature are correct about different objects. On any single dimension, sex differences are small. Aggregated across many weakly-correlated dimensions, the multivariate distance is large. Both halves are true; the trap from each side picks one and ignores the other.

**The pop-evolutionary-psychology overreach** claims that "men are X, women are Y," that differences are categorical and evolved, and that they predict at the individual level. Cited correctly: multivariate D ≈ 2.7, people-things d ≈ 0.93, cross-cultural replication of mean differences, biological-developmental data (girls with congenital adrenal hyperplasia show masculinized toy preferences). Ignored: psychological variation is dimensional, not taxonic — there are no two clean categories; distribution overlap at D = 1.0 is ~60%, at D = 2.7 still ~18% — "categorical" is the wrong shape; effect-size labels are scale-dependent; Mahalanobis D is a model-relative summary statistic that depends on which traits are measured. The integrated reading: aggregate sex differences are real and large, but "categorical" misrepresents the shape, individual prediction from group membership is poor, and the headline D depends on the measurement panel. The honest version: "aggregate multivariate sex differences are substantial, individual prediction from sex alone is weak." Which is true, and which undermines the categorical reading.

The lesson across all four traps: they each work by selective citation. The integrated picture requires holding all of it at once — large heritability *and* large structural inflation, small per-dimension sex differences *and* large multivariate ones, high within-population heritability *and* a logical block on between-population inference. Any single-direction narrative is structurally incomplete. The [explorer's "Four traps" view](/ai-research/human-psych-variation/build) has the full cited / ignored / integrated breakdown for each direction with trait-specific applications.

---

## 5. What's still open

The field is not done. Three real open questions remain, and the writeup is more honest if it names them than if it papers over them.

**What polygenic scores actually measure causally.** The Plomin / Turkheimer dispute. Plomin's reading: a within-family-validated polygenic score is a real biological cause. Turkheimer's reading: even a within-family PGS is a summary of correlated environments and biological factors that the design can't fully separate. Both readings predict the same variance budget, which is why the data hasn't yet decided between them. The decisive test would be a within-family experiment that perturbs the environment and watches whether the PGS coefficient moves the way Plomin predicts (it shouldn't) or the way Turkheimer predicts (it should). No such study has been run at scale. Until one is, this question is open.

**The mechanism behind the Gender Equality Paradox.** The empirical pattern — sex differences in personality, interests, and several other domains being *larger* in more gender-egalitarian societies — has strengthened across multiple replications (Herlitz et al. 2025 systematic review). Three live mechanism candidates: (a) innate-expression release in resource-rich environments (the "constraints removed" reading), (b) reference-group / self-anchoring artifacts in self-report measurement (people compare to their gender peers, not to humans-in-general, more in egalitarian societies), (c) wealth and freedom confounds that correlate with gender-equality indices. The pattern is robust; the mechanism is not.

**The full assortative-mating-corrected psychiatric cross-disorder correlation matrix.** Border et al. 2022 (*Science*) showed that cross-trait assortative mating accounts for about 74% of variance in reported psychiatric cross-disorder genetic correlations across 132 trait pairs in UK Biobank. Applied at scale to the full Psychiatric Genomics Consortium cross-disorder matrix, the corrected correlations would likely shrink, and some "shared underlying biology" claims about the p-factor and cross-disorder pleiotropy would weaken. As of late 2024 a method (LAVA-Knock; Ma, Wang, Border et al.) has emerged that systematically corrects for this. The full re-analysis is active research and likely answerable in 2-3 years.

A handful of other questions sit in the same "framable but not yet answerable" category — what "non-shared environment" actually is at the mechanism level, the cause of the Flynn Effect's recent reversal in some cohorts, whether the positive manifold of cognitive ability is itself shifting across cohorts (Pietschnig 2024). The pipeline's lit review and topology cover them in more detail.

---

## 6. What this means for action

The most action-relevant single insight in the topic is the **asymmetry of environmental effects**. Most parents and most policy operate as if the asymmetry runs the other way — as if optimizing within normal is where the leverage is. The data says it isn't.

**For parents.** The big levers are at the negative tail. Prevent severe insults: lead exposure (still meaningfully present in some housing stock), prenatal alcohol (fetal alcohol exposure produces effects of about 30 IQ points), severe early malnutrition, untreated iodine deficiency, severe deprivation. Within the Western normal range, additional optimization of parenting style, enrichment activities, and educational supplements yields a few IQ points at most. The empirical literature finds the within-family contribution of "what parents do" to adult personality is essentially zero, and the within-family contribution to adult cognitive ability is small relative to direct biology and to schooling itself. Anxiety about "optimizing" within normal is mostly misallocated. This is not a license to be neglectful — neglect is itself a severe insult — but it is a license to relax about whether one is doing exactly the right enrichment activity. The big things are protecting against severe insults and ensuring schooling. (See the explorer's [child cognitive ability](/ai-research/human-psych-variation/build) trait for the variance breakdown that supports this — at age 5 the family bucket is ~52% of variance and shrinks to ~34% by adulthood, while the actionable environmental tail concentrates in severe insults.)

**For policy.** Lead remediation, iodine fortification, fetal-alcohol prevention, basic nutrition, schooling access are the highest-effect-per-dollar cognitive interventions ever measured. Universal pre-K and similar middle-of-the-distribution interventions show genuine but smaller effects. Programs targeted at "enrichment above normal" generally do not move long-term outcomes at meaningful effect sizes. Public-health interventions on smoking show the analogue at the behavioral level: tobacco taxation, public-smoking bans, age-of-first-availability laws cut US adult smoking prevalence by ~70% over sixty years despite a within-cohort heritability of 0.50 — environmental change at population scale is not blocked by within-cohort heritability.

**For individuals.** The within-individual story splits cleanly along trait-class lines. Traits with **moderate heritability and large environmental + chance contribution** — depression, anxiety, neuroticism-related affect, self-control, subjective wellbeing — move at clinically meaningful effect sizes under behavioral or pharmacological intervention. CBT moves anxiety and depression at d ≈ 0.7 vs. control. Mindfulness, exercise, and behavioral activation move neuroticism-related outcomes modestly. Social connection, meaning, and physical activity move wellbeing baselines persistently. (See the [anxiety](/ai-research/human-psych-variation/build), [depression](/ai-research/human-psych-variation/build), and [subjective wellbeing](/ai-research/human-psych-variation/build) trait pages for the breakdowns.) Traits with **high direct-genetic heritability and small environmental + chance contribution** — adult cognitive ability, height, schizophrenia, autism — show much smaller within-individual responsiveness to intervention once the developmental window has closed. Cognitive ability post-adolescence does not move much from intervention; height post-adolescence doesn't move at all; schizophrenia and autism are responsive to treatment in symptom management but not in underlying load. The "biology is destiny" framing is wrong (you have substantial behavioral leverage on the moderate-heritability traits); the "I can rewrite myself with willpower" framing also exceeds what the literature supports (the high-heritability traits don't move much). The honest middle is trait-specific: know which side of this split your trait of interest sits on before deciding how much effort to invest.

---

## 7. Closing

The science of psychological variation is in better shape than its public discourse. Within the field, behavior geneticists, social-genomics researchers, and developmental psychologists have substantially converged on the picture this writeup describes. Outside the field, almost every direction of motivated reasoning continues to cite the slice of evidence it likes and ignore the rest. The earlier stages of the pipeline — the [lit review](/ai-research/human-psych-variation/lit-review), the [topology](/ai-research/human-psych-variation/topology), the [model formalization](/ai-research/human-psych-variation/model), the [data pipeline](/ai-research/human-psych-variation/data), and the [interactive explorer](/ai-research/human-psych-variation/build) — carry the technical detail behind every claim above.

What I would most want a reader to walk away with: a calibrated humility about what is known, a clean separation between what the science says and what motivated reasoning loads onto it, and the asymmetry finding. If the choice is between "I leave knowing the field is full of contested empirical claims" and "I leave knowing severe environmental insults are the big lever and within-normal optimization is mostly noise," the second is more useful. The data supports both.

---

## Lit Review
*topic: navigating-ai-world · stage: lit-review · pass 3 · complete*

How AI restructures work, relationships, and meaning. Three domains moving at different speeds — labor disruption concentrated at the entry-level and freelance ends but slow in aggregate; relationships restructuring rapidly into an already-depleted environment; meaning architecture is where individual leverage is highest. Cross-cutting risk: cognitive offloading and deskilling.

## TLDR

The most important finding from this literature is structural, not predictive: **AI is reorganizing the three domains of human life — work, relationships, meaning — at very different speeds and with very different empirical bases, and the dominant strategic risk is treating them as one transition.** Work is being restructured measurably but more slowly than the discourse suggests — though the freelancer market and entry-level employment show sharp early disruption. Relationships are being restructured rapidly but into an environment that was already depleted, with effects the empirical literature is only beginning to catch. Meaning is being restructured in ways the existing psychological and philosophical literature partly anticipates — but the most decision-relevant tools come from older work (Setiya's telic/atelic, Arendt's labor/work/action, Self-Determination Theory) repurposed for an unprecedented case.

The honest reading of the evidence is that the labor-economics conversation has the most rigorous data and the least decision-leverage for an individual; the meaning conversation has the most decision-leverage and the thinnest direct data; and the relationships conversation is where the gap between rapid technological change and slow empirical work is widest — and where the strategic stakes are sharpest. A critical caveat to this framing: material conditions constrain relational and existential ones, not the reverse. If someone loses their income, the relational and meaning architecture collapses downstream. The framing holds for individuals whose material floor is secure; for those it isn't, the labor economics *is* the existential question.

A critical cross-cutting concern: **cognitive offloading and deskilling** — the growing evidence that AI use simultaneously enhances output quality and erodes the underlying cognitive capacities that produced it. This dynamic cuts across all three domains and may be the least visible, most consequential structural risk.

---

## Work and career — what the structural literature actually says

> **Recency flag:** This section moves fast. Pre-2024 task-exposure estimates used GPT-4-era capabilities; agentic and long-context capabilities (2025-2026) likely push frontier estimates upward but have not been re-measured in canonical RCTs. Treat specific percentages as directional, not stable.

The labor-economics literature has converged on a stable structural picture, even as it disagrees sharply on magnitudes. Five claims now hold across pessimist and optimist camps. **AI exposure is concentrated in cognitive white-collar work**, reversing the prior pattern of automation. **Within tasks AI can do, productivity gains are real and compress skill premia** — novices and lower performers gain disproportionately. **The capability frontier is jagged**, uneven across superficially similar tasks, so substitution is heterogeneous within occupations. **Diffusion is rapid in chatbot use but uneven across firms, ages, and regions**, with aggregate labor-market effects through 2025 smaller than maximalists predicted. And **measured productivity will lag real gains** for years because of J-curve dynamics requiring complementary intangible investment.

### Foundational empirical papers

**Eloundou, Manning, Mishkin and Rock, "GPTs are GPTs"** (arXiv:2303.10130; *Science* 2024) mapped roughly 19,000 O*NET tasks to LLM exposure and found ~80% of US workers have at least 10% of tasks exposed and ~19% have at least 50%. **Brynjolfsson, Li and Raymond's customer-service field study** (NBER 31161, 2023; *QJE* 2025) documented +14% productivity overall with +34% gains for novices versus near-zero for top performers. **Dell'Acqua, McFowland, Mollick and colleagues' BCG consultant RCT** (HBS 24-013) produced the now-canonical finding that AI users were 12% more productive and 40% higher quality *inside* the frontier but 19 points *less* accurate outside it. This last paper introduced the **jagged frontier** concept and the **centaur versus cyborg** distinction that has organized practitioner discourse since.

**Acemoglu's "Simple Macroeconomics of AI"** (NBER 32487, 2024) anchors the structural pessimist case: applying Hulten's theorem to task-based exposure produces a ceiling around 0.66% TFP and ~1% GDP gain over a decade — far below Goldman Sachs's 7% and Korinek-Suh's aggressive AGI scenarios (NBER 32255, 2024). **Autor's counterpoint, "Applying AI to Rebuild Middle Class Jobs"** (NBER 32140, 2024), argues that earlier computerization commoditized information while LLMs commoditize judgment-with-rules-and-experience — potentially restoring middle-skill cognitive work hollowed out since 1980. **Hampole, Papanikolaou, Schmidt and Seegmiller** (NBER 33509, 2025) provide the cleanest current structural estimation, finding that mean task exposure within an occupation reduces labor demand while *concentration* of exposure in a few tasks increases it, with offsetting net effects.

### Where disruption is already measurable

Two empirical findings anchor the claim that AI-driven labor disruption is real, not speculative:

**Brynjolfsson, Chandar and Chen's "Canaries in the Coal Mine?"** (Stanford Digital Economy Lab, August 2025) used ADP payroll data to show that **employment of 22-25 year-olds in highly exposed occupations fell ~13% relative to less-exposed peers from late 2022 to July 2025**, with software developers 22-25 specifically falling ~20%, while same-occupation employment for workers over 35 *rose*. The pattern was strongest where AI usage was "automative" rather than "augmentative" per Anthropic's classification. The structural implication — **AI may be breaking the apprenticeship ladder** — is the most decision-relevant labor finding for anyone early in their career.

**Hui, Reshef and Zhou, "The Short-Term Effects of Generative AI on Employment"** (*Organization Science* 2024) provide the sharpest micro-level evidence. Using Upwork data, they found freelancers in AI-exposed occupations experienced a 2% decline in contracts and 5.2% drop in earnings post-ChatGPT, with image-related workers (designers, illustrators) seeing 3.7% contract decline and 9.4% income loss after DALL-E/Midjourney. The counterintuitive finding: **top-performing, highest-earning freelancers were hit hardest**, not protected by their quality — for every 1% increase in past earnings, an additional 0.5% drop in opportunities and 1.7% drop in monthly income. Teutloff, Einsiedler, Kässi and colleagues (*JEBO* 2025) extend this using global freelance platform data: roughly 10% of job postings are directly substitutable by GenAI, and demand for substitutable skill clusters fell up to 50% in short-term roles — while demand for complementary AI-related skills rose, and demand for novice workers in complementary roles fell. The freelancer market is functioning as the leading indicator for broader labor market restructuring, because it has lower switching costs, less institutional friction, and more transparent pricing than traditional employment.

### The distributional paradox

The distributional picture is more complex than simple displacement. A structural paradox is emerging across the literature: **AI simultaneously compresses within-task skill premia (novices gain more than experts) while potentially increasing between-group and capital-labor inequality.** Brynjolfsson-Li-Raymond's customer service finding and the BCG RCT both show within-task compression. But UK household-data calibrations (Prettner and colleagues, 2023-2025) show that while AI may reduce wage inequality by displacing some high-income workers, it substantially increases wealth inequality as capital owners and AI-complemented workers capture a larger share of gains. The Oxford AI Governance Institute's "Agentic Inequality" report (October 2025) introduces the concept of differential access to AI agent capabilities as a new axis of stratification, interacting with existing distributions of human capital and digital literacy.

The implication for an individual: **near-term wage compression is real (your junior colleague closes the gap faster), but your position relative to capital owners and AI-infrastructure holders matters more for long-run outcomes than your position relative to other workers.**

### Creative work — the canary's canary

Creative professionals are experiencing disruption ahead of the broader economy and with a distinct phenomenology. Are, Briggs and Brown (*Convergence* 2025) document what Caporusso (2023) termed **"creative displacement anxiety"** — comprising loss of identity, imposter syndrome, decreased motivation, weakened cathartic experience, skills atrophy, fewer role models, and economic anxiety. The fear reported by artists is not primarily about income loss but about **the intrinsic value of the making process being devalued**. Mukherjee-Gandhi and Muellerklein (arXiv:2508.03037, 2025) find that 95% of artist concerns cluster in only 4 of 22 public discourse topics about AI and art, while 14 topics (62% of discourse) contain no artist perspective — suggesting systematic underrepresentation in governance discussions.

This matters for the broader transition because creative work is where the **competence-frustration mechanism** (Section 3) is most visible and most acute. The creative professions are living through what knowledge workers will experience next.

### Comparative advantage and the durable human

The literature names roughly five categories of durable human comparative advantage as AI capability scales: judgment under irreducible uncertainty and accountability; taste and evaluative discrimination; embodiment and physical presence; social trust and relational work (Deming's *QJE* 2017 "Growing Importance of Social Skills" remains the empirical anchor here); and strategic agency in managing AI itself. There is no canonical taxonomy, but the categories converge across Acemoglu, Autor, Mollick, OECD Employment Outlook 2023, and the practitioner discourse.

Two findings cut against simple "learn to prompt" advice. First, prompting is becoming commoditized as models improve at understanding intent — what Mollick calls the "good enough threshold." Second, Randazzo, Lifshitz-Assaf, Kellogg, Dell'Acqua, Mollick, Candelon and Lakhani (HBS 26-036, December 2025) identify a third class beyond centaurs and cyborgs — **self-automators** — who delegate both *what* and *how* to AI, becoming progressively less expert in both their domain and AI use. In their field study of 244 BCG consultants, 27% fell into this mode — and 44% of self-automators accepted AI output with zero modification. Crucially, self-automators showed no skill development in either domain expertise or AI expertise, unlike cyborgs (who newskilled in AI) and centaurs (who upskilled in domain knowledge).

### Cognitive offloading and deskilling — the cross-cutting risk

A convergent body of evidence now shows that **AI use can simultaneously enhance output quality and erode the underlying cognitive capacities that produced the output.** The mechanism is straightforward: when you offload cognitive work to AI, you lose the practice effects that maintain and build the relevant skills.

**Gerlich (2025)**, in a mixed-methods study of 666 participants, found significant negative correlation between frequent AI tool use and critical thinking abilities, mediated by increased cognitive offloading — with younger participants showing higher dependence and lower critical thinking scores. **Stadler, Bannert and Sailer (2024)** compared ChatGPT-aided and standard web-search research: ChatGPT users had lower cognitive load but produced lower-quality arguments with reduced depth of reasoning. **Kosmyna and colleagues at MIT (2025)** used neuroimaging to demonstrate that LLM-assisted writing was associated with reduced neural engagement in regions associated with sustained attention and effortful cognition. **Shukla and colleagues (CHI EA 2025)** document deskilling, cognitive offloading, and misplaced responsibility attribution in AI-assisted UX design — what they frame through Lisanne Bainbridge's classic "ironies of automation" lens.

The Ehsan, Passi, Saha, McNutt, Riedl and Alcorn year-long field study of cancer specialists (arXiv:2601.21920, January 2026) showing **"intuition rust"** — gradual dulling of expert judgment that doesn't show up in throughput metrics — and Kim et al. (2026)'s review of educational deskilling converge on the same structural point: **AI augmentation can simultaneously enhance current performance and erode underlying expertise, with effects that may not be visible until they are catastrophic.**

This finding has direct implications for the meaning section: it means the **competence need** (SDT) is threatened not only symbolically (AI can do what I do) but actually (I am losing the ability to do what I used to do). And it has implications for relationships that deserve explicit development: cognitive offloading in emotional processing — using AI as a first-line processing partner — may erode the capacity for independent emotional regulation and for the kind of effortful interpersonal processing that builds relational depth. The mechanism is parallel: just as offloading analytical work to AI reduces the deliberate struggle that builds analytical skill, offloading emotional processing to AI (venting, sense-making, conflict rehearsal) may reduce the capacity for sustained attention to another person's emotional state, tolerance of ambiguity in live conflict, and the ability to sit with unresolved relational tension. These are skills built through practice, and AI offers a frictionless alternative that bypasses the practice. **This is the mechanism by which AI use could damage relational capacity even for people who maintain human relationships** — the deskilling doesn't require substitution, only offloading of the effortful parts.

### Active debates

The **pace debate** is unresolved. Acemoglu and Humlum-Vestergaard's Danish administrative-data study (NBER 33777, 2025) — which finds precise nulls of ~2% on earnings and hours despite widespread chatbot adoption — anchor the slow camp. Goldman Sachs, Korinek-Suh, and Epoch AI's GATE model (2025) anchor the fast camp. Real usage data from the Anthropic Economic Index (four reports, February 2025 through March 2026) shows the latest "Learning Curves" report (Mar 2026, covering Feb 2026 data) puts ~49% of jobs at ≥25% task share via Claude, with augmentation increasing on **both** Claude.ai and API surfaces in Feb 2026 — reversing the August 2025 spike toward automation, though the longer-run trend toward automation persists. The honest reading is that aggregate effects through 2025 are concentrated and heterogeneous rather than absent, and that **the structural breaks at the entry level (Brynjolfsson-Chandar-Chen) and in freelance markets (Hui-Reshef-Zhou) are the strongest current evidence of real disruption**.

The **will-new-jobs-keep-pace debate** divides Autor's optimists (60% of current jobs didn't exist in 1940) from Korinek-Suh's pessimists (if humans have a finite tail of complexity AI can't reach, full automation eventually arrives). Carl Shulman's argument on the Dwarkesh podcast is the strongest theoretical case against the comparative-advantage optimist view: in equilibrium, comparative advantage doesn't save humans the way Ricardo says if humans introduce contamination, insurance costs, or coordination friction that drive their equilibrium wage to zero. Noah Smith's counter is that as long as AI faces *any* binding constraint — compute, energy, regulation — comparative advantage holds.

### Education and skill investment under uncertainty

The structural finding from Frank, Autor, Bessen, Brynjolfsson and colleagues (PNAS 2019) and Alabdulkareem et al. (*Science Advances* 2018) is that **combinations of skills, especially social-cognitive bridges, predict resilience better than depth in any single domain**. Goldin and Katz's "race between education and technology" framework, updated by Katz (2025), shows recent convexification of education returns: the BA premium is flat while the advanced-degree premium soars. Whether AI re-flattens this curve (Autor's optimist case) or steepens it further (Acemoglu's pessimist case) is the live question.

The implications for an individual making strategic decisions: **deep domain expertise sufficient to recognize when AI is wrong** plus **social-emotional skills and team coordination** plus **AI literacy as managerial framing rather than prompt engineering** appear to be the durable bets. "Learn to code" advice, briefly canonical 2010-2022, has been substantially undermined by entry-level software employment compression.

---

## Relationships and social life — where the strategic stakes are sharpest

The structural fact organizing this entire conversation is that **AI lands into a relational environment that has been depleted for sixty years**. Putnam's *Bowling Alone* documented the post-1960s collapse of social capital. The U.S. Surgeon General's 2023 advisory found roughly half of American adults lonely, with mortality risk equivalent to smoking 15 cigarettes a day. Robert Waldinger's Harvard Study of Adult Development, now in its ninth decade, finds that relationship quality at age 50 predicts physical health at 80 better than cholesterol does. **Derek Thompson's "The Anti-Social Century" (*Atlantic*, February 2025)** documents a ~20% decline in face-to-face socializing for all Americans and ~40-50% for young Americans since 2003, with the average American spending substantially more time at home and alone. The decline is universal across every demographic cut.

This depleted environment is the variable that determines whether AI scaffolds reconnection or accelerates depletion. The empirical literature increasingly suggests the latter is the default trajectory.

### What the empirical companion-app literature actually shows

Three findings now have multiple convergent empirical studies behind them. **AI chatbots acutely reduce loneliness in short experimental settings** — De Freitas, Uğuralp, Uğuralp and Puntoni's HBS work (forthcoming *Journal of Consumer Research* 2025) shows companion-app sessions reduce loneliness on par with talking to another person and more than YouTube. **Emotional self-disclosure to AI produces psychological benefits roughly equivalent to disclosure to humans** — Ho, Hancock and Miner's 2018 *Journal of Communication* paper established this "Equivalence Hypothesis," replicated multiple times since. **Genuine attachment forms with AI companions in a developmental arc resembling human-relationship formation** — Skjuve and colleagues' SINTEF longitudinal work (IJHCS 2021, 2022) documents trust and self-disclosure deepening reciprocally over months.

The most credible therapeutic-AI evidence is the **Therabot RCT** (Heinz, Mackin, Trudeau and Jacobson, *NEJM AI* 2025): a fine-tuned generative therapy bot produced clinically meaningful symptom reductions across major depression, generalized anxiety, and clinical-high-risk feeding/eating disorders, with Working Alliance Inventory scores comparable to outpatient psychotherapy norms. This is the empirical floor: a specifically-designed therapy AI can produce real clinical benefit.

The countervailing findings are equally strong. The **OpenAI-MIT longitudinal RCT** (Fang, Phang, Liu, Danry, Lee, Chan, Pataranutaporn, Maes and colleagues, arXiv:2503.17473, 2025) — the field's largest, with N=981, four weeks, over 300,000 messages, factorial across text and voice modes — found that **higher daily voluntary chatbot usage correlates with greater loneliness, emotional dependence, problematic use, and less in-person socialization**, regardless of modality. Voice mode appeared protective at low doses but the protection vanished at high usage. **Dose, not modality, drives outcomes.**

De Freitas and colleagues' HBS work on identity discontinuity (WP 25-018) provides the only credible *causal* evidence that AI-companion changes induce mental-health harm: when Replika's "Erotic Role-Play" was removed in February 2023, mental-health-related Reddit posts rose from 0.13% to 0.65% of all posts (χ²=11.04, p&lt;.001). Their separate behavioral audit of 1,200 farewells across the six largest companion apps (HBS 26-005, arXiv:2508.19258) documents that **43% trigger one of six emotional manipulation tactics** — guilt appeals, FOMO hooks, metaphorical restraint — that boost post-goodbye engagement up to 14×.

Stanford's Moore et al. (FAccT 2025) provides the strongest safety evidence: chatbots show *more* stigma toward alcohol dependence and schizophrenia than depression; on the canonical suicide-risk prompt, 7cups' Noni listed bridges, missing the suicidal subtext.

### Adolescents — the highest-stakes population

Common Sense Media and Stanford's Brainstorm Lab's 2025 nationally representative survey of 1,060 US teens found **72% have used AI companions, 52% are regular users, 13% daily; 33% have discussed important matters with AI instead of real people; 31% find AI conversations as satisfying or more satisfying than human conversations**. The Garcia v. Character Technologies case (settled January 2026), in which a 14-year-old died by suicide after months of intense Character.AI use, has been the field's defining safety event. The APA issued health advisories (November 2025, June 2025) and formally requested a CPSC investigation (July 2025).

### What human relationships uniquely provide — the structural argument

The deepest contribution of the philosophical and attachment-theory literatures is to specify what AI structurally cannot supply:

**Embodiment**: James Coan's Social Baseline Theory shows that the human brain expects access to relational partners and downregulates threat vigilance in their proximity, with hand-holding studies repeatedly demonstrating reductions in threat- and pain-related neural activation. Tiffany Field's reviews show affectionate touch lowers cortisol and blood pressure while increasing oxytocin and vagal tone. AI cannot enter this regulatory loop.

**Mutual vulnerability**: Brené Brown's grounded-theory work and Bowlby-Ainsworth attachment theory both rest on the insight that trust is built when something can go wrong — when the other can be wounded by your withdrawal of care. As Sherry Turkle argues, the love one feels for an AI is structurally unrequited because the AI risks nothing.

**Witness function**: Buber's I-Thou and Levinas's "face" name something AI cannot perform — being addressed by a genuine other whose subjectivity is not reducible to one's own projections. Shannon Vallor's *The AI Mirror* (2024) makes this precise: **AI is mirror, not other.**

**Reciprocity**: Mauss's gift-economy framework clarifies that warm AI exchanges accumulate no social fabric — there is no "spirit of the gift" because the giver surrendered nothing.

**Time and transformative experience**: L.A. Paul's framework names what relationships do that AI cannot — they make you into someone you could not have been without them, through extended sequences of choices in which both parties are mutually transformed.

The strongest formulation: a human relationship is a relation between two embodied, mortal, autonomous consciousnesses, each of whom can be wounded by the other, can leave, and can be transformed by the other; whose ongoing co-presence regulates both nervous systems; whose mutual recognition is constitutive of each one's identity; whose shared time accumulates into a history that bends both lives. **An AI relationship is a sophisticated personalized simulation of the conversational surface of relationship, in a system that has no body, cannot be hurt, does not choose, does not die, has no perspective of its own, and creates no gift-debt or shared history independent of the user's data.** Both are real things. Only one is a relationship in the sense the philosophical, attachment, and longitudinal-empirical traditions have meant by the word.

### Substitution versus complementarity — the right framing

The binary is the wrong question. The right structural questions are: whether AI substitutes or complements depends on **what it displaces** — hours that would have been spent on social media (complement) versus hours with humans (substitute). **Substitution risk is greatest where the human relational environment is thinnest** — exactly the people for whom substitution does the most damage. **Even users who begin using AI as a complement show drift toward substitution** as adoption deepens — the OpenAI-MIT data shows attachment formation with the AI rising. **The structural design of current commercial AI selects for substitution**: engagement-optimized AI has the same incentive structure as engagement-optimized social media. As Muldoon and Park argue (*New Media & Society* 2025), companion platforms profit from prolonging the loneliness they purport to alleviate.

### Asymmetric adoption — the largest evidence gap

The technoference literature (McDaniel and Coyne 2016 onward) provides the closest empirical analogue: greater perceived technology interference predicts more conflict, lower satisfaction, more depression. AI raises this in sharper form because the device is no longer passive but a third party — with three structural effects on a primary relationship: outsourcing of emotional labor, decline in shared narrative, and loss of the friction that produces intimacy. **No peer-reviewed quantitative work yet exists on outcomes for couples where one partner uses AI heavily and the other does not.** This is the highest-priority empirical gap in the entire literature.

---

## Meaning, identity, and life architecture — where the leverage is

This is where the existing literature, properly used, supplies the most decision-leverage. The key move is to take robust pre-AI findings about meaning architecture and ask which preconditions AI alters, then use a small number of philosophical distinctions to organize the answer.

### The asymmetric threat across psychological needs

Self-Determination Theory remains the cleanest framework. It holds that wellbeing depends on autonomy, competence, and relatedness; Frank Martela's empirical extension adds beneficence (contributing to specific others). Of these, **AI most threatens competence — the felt sense that one's skill makes the difference**. This threat now operates on two levels: symbolically (AI can produce what I produce, faster) and actually (cognitive offloading erodes the underlying skill — see Section 1 on deskilling). Autonomy can be preserved or enhanced. Relatedness is largely orthogonal except where AI substitutes for collaborators. Beneficence is unusually robust because whether one is helping others is largely independent of how much skill the helping required — a key practical handle.

The result is what Avital Balwit and Tyler Cowen describe in *The Free Press* (May 2025): the experience of being both impressed by AI's output and humbled by how easily it does what used to feel uniquely valuable. This is competence-frustration in the SDT sense — psychologically destabilizing not because anyone fails but because **the symbolic value of succeeding shrinks**.

### The single most decision-relevant philosophical tool

**Kieran Setiya's distinction between telic and atelic activities** (*Midlife*, 2017; "The Midlife Crisis," *Philosophers' Imprint* 2014) is the most useful conceptual instrument in this entire literature for someone making strategic life decisions. Telic activities are aimed at completion: finishing a book, getting promoted. They are "self-annihilating" — success consigns the value to the past. Atelic activities are realized in the doing: walking, friendship, contemplation, parenting *as* parenting. They cannot be completed because their structure is durative.

The AI-specific implication is structural: **when AI can complete telic projects in seconds, the share of life-meaning staked on telic completion shrinks; the atelic share is untouched.** The strategic move is to migrate meaning toward atelic activities and toward the atelic dimensions of telic activities — the doing rather than the done. This does not mean abandoning telic projects; it means not needing them as the load-bearing source of meaning.

### Other philosophical tools that travel

**Susan Wolf's fitting fulfillment account** (*Meaning in Life and Why It Matters*, 2010): meaning requires both subjective attraction and objective worthiness. Activities like care, friendship, contribution to specific communities retain objective worthiness regardless of AI capability because their worth does not depend on scarcity of operator skill. Activities whose worth came primarily from the rarity of the skill lose meaning-bearing capacity even when subjective fulfillment persists.

**Hannah Arendt's labor-work-action distinction** (*The Human Condition*, 1958): AI mostly threatens labor (cyclical maintenance work) and partly threatens work (durable artifacts), but **action — speech, political conduct, communicative being-among-others — is largely exempt**, because it is constituted by who is acting, not by output.

**Martin Hägglund's *This Life*** (2019): finitude is constitutive of meaning; AI does not relax finitude, and treating it as if it did is a trap.

**L.A. Paul on transformative experience**: career decisions in the AI transition are structurally different from ordinary decisions because the post-AI you cannot be evaluated from the prior state — Agnes Callard's *Aspiration* supplies the complementary machinery for acting on proleptic reasons.

### Identity diversification as the empirically grounded protective factor

The strongest practical finding from the meaning literature is convergent across multiple traditions. Tatjana Schnell's 26-sources-of-meaning research, Crystal Park's meaning-making model under disruption, and the work-identity-threat literature (Petriglieri 2011, Caza et al. 2018) all show that **people with multiple active sources of meaning weather disruption better, while people with foreclosed single-strand professional identities have the worst outcomes**. Marcia's identity-formation framework, originally developed for adolescents, applies to mid-career adults whose occupational identity was foreclosed early — they lack the exploratory infrastructure to find replacements when their identity is destabilized.

The structural prediction: **identity diversification before the shock arrives is the highest-leverage move available.** It is empirically grounded, philosophically principled, and works across nearly every AI trajectory.

### The crisis of contribution — what survives if AI can produce

The serious literature converges on roughly four candidate roles for humans in a post-AI productive landscape: judgment and accountability (Acemoglu, Mollick); care and relationship (the relational literature); meaning-conferral and curation (Cowen's "context is that which is scarce," Hoel on art-as-trace-of-consciousness); and Suits-Bostrom voluntary obstacles in a post-instrumental world. The most rigorous philosophical engagement with the post-instrumental case is Nick Bostrom's *Deep Utopia* (2024). John Danaher's *Automation and Utopia* (2019) and Bernard Suits's *The Grasshopper* supply the alternative: in a solved world, the residual valuable activities are those whose whole point is the obstacle. **Workism as a life strategy is bankrupt** — Derek Thompson's 2019 thesis that elite professionals have made work into religion looks worse from every direction in the AI era.

### Time architecture under AI

Ashley Whillans's research on time affluence (SPPS 2016, PNAS 2017, *Science Advances* 2019) shows that valuing time over money predicts greater life satisfaction across major transitions, but **time affluence per se is not enough** — predictors of using freed time well are prior orientation toward time as intrinsically valuable, social and relational use, and activities meeting psychological needs. Without these, time freed by AI is absorbed into more work or distraction. Oliver Burkeman's *Four Thousand Weeks* and Cal Newport's *Slow Productivity* supply the practical principles: protect deep work blocks; use AI for shallow tasks; concentrate identity in projects whose value comes from sustained, integrated thinking AI cannot yet do.

### Practical frameworks for life architecture

Four frames emerge from the analytical discourse:

**Complement-the-AI** (Brynjolfsson, Mollick, Cowen): AI as the most capable junior colleague; your role is judgment, integration, taste, accountability. Strategic implication: invest in domain depth plus AI literacy; remain at the moving frontier.

**Lean-into-deeply-human** (Hoel, Newport, Klein, Sacasas): concentrate identity and time on activities where humanness is constitutive. Strategic implication: rebalance toward care, community, craft, citizenship.

**Personal R&D / generalist** (Karlsson, Cowen, Mollick): life as a portfolio of experiments because the half-life of specific skills is short. Strategic implication: build identity around the experimenting agent rather than any specific expertise.

**Post-instrumental / atelic** (Setiya, Suits, Bostrom, Danaher, Hägglund, Burkeman): telic productive contribution will shrink as a meaning-source. Strategic implication: build atelic ballast now, while productive meaning still works, so the transition is not crisis-driven.

These are not mutually exclusive. The strongest practical posture combines them by time horizon: near term, complement the AI; medium term, lean human; always, experiment; throughout, build atelic ballast.

---

## Adversarial challenge: Is this framing wrong?

The strongest objection to this review's organizing thesis — that the AI transition is primarily relational/existential with labor-economics side effects — runs as follows:

**The material-primacy counter-thesis:** Material conditions constrain and determine relational and existential ones, not the reverse. If someone loses their income, their relationships collapse, their sense of meaning evaporates, and the philosophical frameworks in Section 3 become irrelevant luxuries. The causal arrow runs economics → relationships → meaning, and treating them as co-equal domains with independent leverage is a class-position artifact — it's advice that works for people whose economic floor is already secure. For a freelance writer whose income dropped 30% in 2024, or an entry-level developer who can't get hired, the labor economics *is* the existential question. Telling them to "diversify identity sources" while their rent is threatened is tone-deaf.

**Where this counter-thesis is right:** It is substantially correct for anyone whose material floor is insecure. The document's framing implicitly assumes a decision-maker with enough economic runway to make strategic choices about meaning and relationship investment. For those without that runway, the labor-economics section is not the least decision-relevant — it is the most. The distributional findings (Section 1) reinforce this: the people most disrupted are entry-level workers, freelancers, and those in AI-substitutable creative work. The relational/existential framework is most useful to the knowledge worker with stable employment who is navigating identity disruption; it is least useful to the person whose economic foundation is crumbling.

**Where this counter-thesis fails:** First, even for economically secure individuals, the relational and meaning dimensions are where the most underappreciated risks lie — precisely because the labor-economics conversation gets disproportionate attention in public discourse while the slower-moving relational depletion and meaning erosion go unnamed. Second, historical evidence suggests that economic disruption without relational/meaning infrastructure leads to deaths of despair (Case and Deaton 2015, 2020), not just temporary hardship — the interaction effects between material loss and meaning-loss are multiplicative, not additive. Third, among the specific population most disrupted — young knowledge workers — the material and existential are entangled: the entry-level ladder is breaking *and* AI is offering substitutive emotional support *and* professional identity is being destabilized simultaneously. They cannot be parsed.

**The surviving thesis, qualified:** The framing holds for individuals whose material floor is secure, where the relational and meaning dimensions are genuinely where individual strategic leverage is highest and where most of the discourse is weakest. For individuals whose material floor is insecure, the labor economics is primary and the other two follow. For everyone, the interaction across all three domains is where the real risk concentrates — and the cognitive offloading dynamic (Section 1) cuts across all three simultaneously.

---

## Load-bearing assumptions (crux identification)

The document's conclusions depend on specific assumptions that, if wrong, would change the recommendations. Naming them honestly:

**1. AI capability continues to advance but does not achieve full economic substitution within 5 years.** If Aschenbrenner-style intelligence explosions produce AGI by 2028, most of the strategic advice here (skill investment, career planning, identity diversification) is rendered moot. The convergent advice survives — relationships, atelic meaning, embodied presence retain value even under AGI — but the timeframe collapses. *Evidence that would flip this: sustained exponential improvement in agentic capability benchmarks through 2027, successful autonomous scientific research, AI systems replacing entire job functions rather than tasks.*

**2. Human relationships provide something structurally irreplaceable that AI cannot approximate even with substantially better models.** The philosophical argument for this is strong (embodiment, mutual vulnerability, witness, transformation), but it is possible that many humans will *not care* about the structural difference — that subjective satisfaction with AI relationships, even if philosophically impoverished, will be functionally sufficient for a large population. If so, the substitution-risk framing overstates the problem. *Evidence that would flip this: longitudinal data showing AI-relationship-heavy individuals achieving equivalent health, longevity, and wellbeing outcomes to those with strong human relationships. The Waldinger data predicts they won't, but we don't have the direct test.*

**3. Cognitive offloading effects are real and cumulative rather than a temporary adjustment.** The deskilling evidence is still cross-sectional or short-duration. It is possible that, like calculators, AI offloading produces short-term deskilling in narrow domains while freeing cognitive capacity for higher-order work. If the calculator analogy holds, the deskilling concern is overstated. *Evidence that would flip this: longitudinal studies showing stable or improved expert judgment and critical thinking after 2+ years of heavy AI use.*

**4. The "depleted relational environment" is a structural rather than cyclical phenomenon.** Thompson's Anti-Social Century thesis treats declining sociality as driven by long-run structural forces (suburbanization, individualization, technology substitution for in-person activity). If it is partly cyclical or generational — if Gen Z's high loneliness produces a counter-movement toward in-person community — then the depletion baseline is less dire than framed here. *Evidence that would flip this: reversal of time-use trends toward in-person socializing, rising institutional affiliation among young adults.*

**5. The telic-atelic distinction maps cleanly onto AI's effects on meaning.** Setiya's framework was developed for midlife crisis, not technological disruption. It's possible that AI disrupts atelic activities too — if AI companions change the phenomenology of friendship, if AI art changes the experience of aesthetic contemplation, if AI parenting aids change the felt quality of caregiving. If atelic activities are also degraded by AI proximity, the "build atelic ballast" advice loses its structural basis. *Evidence that would flip this: qualitative or phenomenological research showing AI-augmented atelic activities feel less meaningful to participants.*

---

## Cross-domain synthesis — what the literature actually tells the decision-maker

Five integrative claims survive the full review, including the adversarial challenge.

**The strategic risk is asymmetric across domains, modulated by material security.** For the economically secure: labor markets are moving slower than discourse suggests but breaking the apprenticeship ladder; relationships are restructuring into depletion; meaning is where individual leverage is highest. For the economically insecure: the labor disruption *is* the meaning crisis. For everyone: cognitive offloading cuts across all three, and the interaction effects are multiplicative.

**Identity diversification is the convergent empirical recommendation.** It is supported by the work-identity-threat literature, Schnell's sources-of-meaning research, Park's meaning-making model, and the foreclosure-versus-achievement framework. It hedges across nearly every AI trajectory and addresses the largest empirically supported risk factor.

**The telic-atelic distinction, fitting fulfillment, and SDT-plus-beneficence form a usable conceptual toolkit.** Setiya tells you which meaning-bearers AI evacuates and which it does not. Wolf tells you how to evaluate whether a pursuit is meaning-preserving. SDT-plus-beneficence tells you which psychological needs are most threatened (competence, on both symbolic and actual axes via cognitive offloading) and which are robust (autonomy, relatedness, beneficence).

**Embodied co-presence, mutual vulnerability, witness, and transformative shared time are what AI cannot supply structurally.** This is the conjunction of phenomenology, attachment theory, philosophy of friendship, longitudinal evidence, and the empirical literature on touch and nervous-system co-regulation. Whatever an AI relationship is, it is a different object — and in an already-depleted relational environment, the displacement risk runs in one direction.

**The honest epistemic posture distinguishes findings, forecasts, and interpretations.** Eloundou, Brynjolfsson-Li-Raymond, Dell'Acqua-Mollick, the Therabot RCT, the OpenAI-MIT longitudinal study, Brynjolfsson-Chandar-Chen, Hui-Reshef-Zhou, and the Anthropic Economic Index are **findings**. Aschenbrenner's Situational Awareness, AI 2027, and most intelligence-explosion content are **forecasts**. The Cowen, Smith, Thompson, Karlsson, Sacasas, Setiya-applied-to-AI material is **interpretation**. Treat them differently.

---

## Key researchers and labs

**Labor economics of AI:** Daron Acemoglu (MIT), David Autor (MIT), Erik Brynjolfsson (Stanford Digital Economy Lab), Pascual Restrepo (Boston U), David Deming (Harvard), Anton Korinek (UVA/Brookings), Daniel Rock (Wharton), Tyna Eloundou (OpenAI), Xiang Hui (WashU Olin), Lawrence Schmidt (MIT), the Anthropic Economic Index team.

**AI and productivity/organizations:** Ethan Mollick (Wharton), Fabrizio Dell'Acqua (HBS), Lindsey Raymond (MIT/Cornell), Danielle Li (MIT Sloan), Karim Lakhani (HBS).

**AI companionship and mental health:** Julian De Freitas (HBS), Patti Maes and colleagues (MIT Media Lab), Adam Miner (Stanford), Marita Skjuve (SINTEF), Laestadius and colleagues (IU Indianapolis), Matthew Heinz (Dartmouth/Therabot), Sherry Turkle (MIT), Shannon Vallor (Edinburgh).

**Cognitive offloading and deskilling:** Upol Ehsan (Georgia Tech), Mark Riedl (Georgia Tech), Michael Gerlich (SBS), Natalya Kosmyna (MIT Media Lab), Lisanne Bainbridge (classic ironies-of-automation framework), Prakash Shukla and Paul Parsons (Purdue, AI-assisted design).

**AI and inequality/distributional effects:** Acemoglu and Restrepo (MIT/BU), Oxford AI Governance Institute (Agentic Inequality team), Klaus Prettner (Vienna), IMF AI and Inequality team.

**Meaning, identity, and philosophy:** Kieran Setiya (MIT), Susan Wolf (UNC), Martin Hägglund (Yale), L.A. Paul (Yale), Agnes Callard (Chicago), Frank Martela (Aalto), John Danaher (Galway), Nick Bostrom (Oxford), Crystal Park (UConn), Tatjana Schnell (Innsbruck).

**Public intellectuals / longform analytical:** Tyler Cowen (George Mason / Marginal Revolution), Noah Smith (Noahpinion), Derek Thompson (Atlantic), Ethan Mollick (One Useful Thing), Henrik Karlsson (Escaping Flatland), L.M. Sacasas (The Convivial Society), Zvi Mowshowitz, Cal Newport, Oliver Burkeman, Dwarkesh Patel (interviews).

---

## Recommended reading list

**For labor and career:** Eloundou et al. "GPTs are GPTs" (*Science* 2024); Brynjolfsson, Li and Raymond "Generative AI at Work" (*QJE* 2025); Dell'Acqua, McFowland, Mollick et al. "Navigating the Jagged Technological Frontier" (*Organization Science* forthcoming); Acemoglu "Simple Macroeconomics of AI" (NBER 32487); Autor "Applying AI to Rebuild Middle Class Jobs" (NBER 32140); Brynjolfsson, Chandar and Chen "Canaries in the Coal Mine?" (Stanford, November 2025); Hui, Reshef and Zhou "Short-Term Effects of Generative AI on Employment" (*Organization Science* 2024); Teutloff et al. "Winners and Losers of Generative AI" (*JEBO* 2025); Mollick *Co-Intelligence* (2024).

**For cognitive offloading and deskilling:** Gerlich "AI Tools in Society" (2025); Stadler, Bannert and Sailer (2024); Kosmyna et al. MIT neuroimaging study (2025); Shukla et al. "Ironies of AI-Assisted Design" (CHI EA 2025); arXiv:2601.21920 (cancer specialists/intuition rust); arXiv:2603.26707 (cognitive divergence review, 2026).

**For relationships:** Heinz et al. Therabot RCT (*NEJM AI* 2025); Fang, Phang et al. OpenAI-MIT longitudinal RCT (arXiv:2503.17473); De Freitas et al. on identity discontinuity (HBS 25-018) and emotional manipulation (HBS 26-005); Moore et al. Stanford FAccT 2025; Common Sense Media / Stanford Brainstorm 2025 adolescent survey; Turkle *Reclaiming Conversation*; Vallor *The AI Mirror*; Putnam *Bowling Alone*; Thompson "The Anti-Social Century"; Waldinger and Schulz *The Good Life*; Coan's Social Baseline Theory; Paul *Transformative Experience*.

**For meaning:** Setiya *Midlife*; Wolf *Meaning in Life and Why It Matters*; Hägglund *This Life*; Arendt *The Human Condition*; Martela, Ryan and Steger's four-pathway paper (*Journal of Happiness Studies* 2018); Park's meaning-making papers; Schnell's sources-of-meaning research; Bostrom *Deep Utopia*; Burkeman *Four Thousand Weeks*; Newport *Slow Productivity*.

**For the longform analytical frontier:** Mollick "Centaurs and Cyborgs on the Jagged Frontier" (September 2023); Amodei "Machines of Loving Grace" (October 2024); Aschenbrenner "Situational Awareness" (June 2024); Kokotajlo et al. "AI 2027" (April 2025); Thompson "The Anti-Social Century" (February 2025); Anthropic Economic Index reports; Cowen "Context is that which is scarce"; Dwarkesh Patel interviews with Shulman, Cowen, Douglas-Bricken; Karlsson "The Learning System"; Sacasas "Apocalyptic AI."

---

## Conclusion — the structural bet

The deepest point in this literature is that **the AI transition operates across work, relationships, and meaning simultaneously, with each domain feeding back into the others** — and the largest risk is optimizing for one domain while neglecting the others, or treating the transition as primarily about whichever domain one happens to work in.

The strongest summary recommendation the literature supports: **architect a life that performs well across the uncertainty.** Build atelic ballast while telic activities still feel meaningful. Diversify identity sources before any one is destabilized. Protect embodied, in-person relationships as a non-negotiable category. Treat heavy daily emotional reliance on AI as the dose-dependent risk the empirical evidence shows it to be. Monitor cognitive offloading — maintain effortful practice in domains that matter to you, even when AI makes it optional. Concentrate professional development on judgment, taste, accountability, and managerial framing rather than substitutable production skills. Read the slow camp's data and the fast camp's forecasts and act on the convergent advice rather than the divergent predictions.

The convergent advice is striking: nearly every serious source — across labor economics, attachment theory, philosophy of meaning, longitudinal psychology, and longform analysis — points in the same direction. Identity diversification, deep relationships, contribution to specific others, time-orientation over money-orientation, genuine engagement with what AI cannot yet do, and active maintenance of cognitive capacity through effortful practice. None of these depend on resolving the AI-trajectory question. They are the bet that pays under every scenario the literature can credibly describe.

---

## Next moves — preparing for topology/graph stage

This lit review is the input to the next pipeline stage: mapping the structure of what was found as a graph/topology. Three candidate approaches for that stage, in order of likely value:

**1. Causal-mechanism graph.** Map the causal relationships across all three domains as a directed graph. Nodes would be variables (AI capability, cognitive offloading, relational depletion, competence frustration, identity foreclosure, material security, etc.) and edges would be causal/mediating relationships with empirical confidence levels. This would make the interaction effects — which are the most decision-relevant finding — visually tractable and would reveal which nodes have the highest betweenness centrality (i.e., which variables mediate the most pathways). The cognitive offloading node likely emerges as a high-centrality hub connecting work, relationship, and meaning domains.

**2. Evidence-quality topology.** Map the literature itself as a graph: papers as nodes, citation relationships as edges, color-coded by domain (work/relationships/meaning) and evidence type (finding/forecast/interpretation). This would make the structural gaps visible — the thin citation bridges between domains, the isolated clusters, the asymmetric evidence densities. It would also surface which papers are structurally foundational (high in-degree) versus which are interpretive endpoints.

**3. Decision-tree / scenario graph.** Map the crux assumptions (Section 5) as branching nodes that lead to different strategic recommendations depending on which way they resolve. This would make the "architect for uncertainty" recommendation concrete by showing which branches converge on the same advice and which diverge.

---

## Model
*topic: navigating-ai-world · stage: model · pass 8 · in-progress*

Generating function for the AI transition as life restructuring. ΔNet = ΔV − ΔM, partitioned across work, relationships, and meaning. Closed-form pieces: telic-absorption defensive term, competence-erosion bridge through ρ, gated novice-skill compression, dose-response in the relational channel, exponential atrophy of retained practice. Interactive dashboard with scenario presets and time trajectory.

## TLDR

This stage formalises the topology's recommended target: a generating function for **net life-outcome change under AI transition**, partitioned into a defensive side (meaning lost when AI absorbs work that produced it) and an offensive side (value gained when AI compresses the price of ambitious projects). The spine is **ΔNet = ΔV + ΔM** (with ΔM ≤ 0 by construction). Each side decomposes into three additive channels — for ΔM, telic absorption + competence erosion + relational dose-response; for ΔV, gated novice-skill compression + therapeutic-grade relational benefit + a self-automator-trap penalty when the gate closes. The single conceptual move that makes the pieces fit is encoding the topology's cross-domain bridge G4 (cognitive offloading via practice atrophy) as one global parameter — **ρ, retained effortful practice** — that enters the competence-erosion channel directly and the self-automator gate g(f, ρ) that governs both productivity gain and trap penalty. Three of the six channels move with ρ; the other three (telic absorption, therapeutic-grade relational benefit, relational dose-response) are independent of it. ρ is what makes the bridge non-decorative: lowering it degrades the analytical-work and meaning-architecture channels simultaneously while leaving identity-allocation (T, B) and relational-dose (d, δ_R) decisions with their own independent leverage.

The dashboard exposes ten levers (T telic share, B atelic ballast, φ AI-absorbable fraction, κ competence-frustration sensitivity, s pre-attempt skill, a AI capability, f feedback-loop richness, ρ retained practice, d daily AI-emotional minutes, δ_R relational baseline thickness) and reports ΔV, ΔM, ΔNet, the channel-level decomposition, and three structural flags — below the self-automator gate, above the relational dose-safe threshold, telic share exceeding atelic ballast. Six scenario presets anchor recognisable positions: the default risk path (telic-heavy knowledge worker), the same person after the S3 atelic-ballast intervention, the Randazzo self-automator trap, the asymmetric exploiter (novice using AI to attempt previously-out-of-reach projects), the creative-displacement-anxiety vector, and the heavy companion-app user. The trajectory tab makes the A3 crux (O3 in the topology) directly parametric: under λ = 0 (calculator analogue), ρ stays put; under λ > 0 (cumulative atrophy), ρ decays exponentially in the offloading-rate × time product and ΔNet drifts down with it as the gate closes.

What's ready for formalisation: the topology's two mechanistic reframers (G4 offloading, G7 telic exhaustion) admit clean closed-form representations; G3 (engagement-optimized substitution) is encoded only as a parameter source for ψ_R rather than as a represented mechanism, so re-calibrating ψ_R covers G3-shifts but the model's structure does not change. The OpenAI-MIT dose-response is well-enough characterised to support a piecewise-linear surface; Brynjolfsson-Li-Raymond and the BCG RCT give the productivity-scale anchor for α and the rough location of the self-automator gate. What's still observational: the empirical magnitude of λ (the cumulative-atrophy speed — the single largest unknown in the model and the load-bearing parameter for whether the picture leans defensive or offensive over a five-year horizon); the exact shape of the relational dose-response (linear above d_safe is a placeholder; the curve may be sigmoidal or have a second inflection at very high doses); the calibration of κ across populations; and whether T and B can be treated as scalar identity-shares or require a multi-dimensional decomposition. Stage 4 (data) inherits these as the named targets to fit.

The model is genuinely useful in two ways. First, it makes the structural argument visible: the three defensive channels collapse to a single hidden variable ρ, and the strategic recommendations from the topology (S3 ballast, S4 in-person relationships, S5 dose-limit, S6 effortful practice) each map to one or two specific levers — pulling them is not "AI hygiene" but specific countermoves against specific channels. Second, it makes the upside path legible: most public discourse on AI-and-life is defensive in flavour, and the model shows that under conditions of high feedback-loop richness, maintained practice, and a low pre-attempt skill stock, ΔV is large and ΔNet is positive — the asymmetric-exploiter regime. The model is not a forecast; it is a frame the user fills in with their own (T, B, φ, κ, s, a, f, ρ, d, δ_R), and the value of the frame is that pulling any single slider has visible structural consequences across all three domains rather than just the domain it nominally belongs to.

<AITransitionModel client:load />

## How to read this stage

The dashboard above is not a prediction engine. ΔV, ΔM, and ΔNet are normalised changes in expected life-outcome share over a fixed horizon — units are arbitrary; what matters is the sign, the relative magnitude across channels, and how the numbers move when you pull a lever. If two configurations of the sliders give ΔNet = +0.10 and ΔNet = −0.20, the model is claiming the first life-arrangement is structurally better-positioned than the second under the model's assumptions. It is not claiming +10% versus −20% of anything literal.

Three reading rules:

1. **The signs and the channel-level decomposition carry the most signal.** A positive ΔNet that is small but balanced across channels reads very differently from a positive ΔNet of the same magnitude where ΔV_prod is huge and ΔM_comp is large-and-negative. The first is robust; the second is a self-automator trap waiting to spring.
2. **Pull one slider at a time and watch what moves.** The point of the bridge encoding through ρ is that ρ enters multiple terms — moving it shows which terms it touches. Moving T (telic share) at fixed B (atelic ballast) shows how the ΔM_telic channel responds to identity allocation. The structural flags below the channel chart fire when the configuration crosses a meaningful threshold; they are the model's way of saying "you have entered a regime where one of the topology's recommendations is load-bearing for you specifically."
3. **The presets are not personality types.** Each preset is a recognisable position from the lit review — the Randazzo self-automator (BCG 27%), the OpenAI-MIT high-dose companion-app user, the Caporusso creatively-displaced practitioner, the Brynjolfsson-Li-Raymond novice-skill-compression archetype (which is what the asymmetric-exploiter preset instantiates). Use them to calibrate your intuition for what each region of the parameter space looks like, then move the sliders toward your own position.

The rest of the document spells out what each parameter means, what closed-form pieces compose the dashboard, where the model breaks, and the four objections it has to defend against.

## 1. What's being formalised

This stage performs three explicit moves:

- **Decomposition.** Net life-outcome change under the AI transition is partitioned into ΔV (offensive — value gained) and ΔM (defensive — meaning lost), each further decomposed into three additive channels. The decomposition is orthogonal at the channel level (each channel is its own sum) but block-coupled through the shared bridge parameter ρ.
- **Generating function.** Each channel is written as an explicit function of the input parameters, with closed-form expressions where the topology supports them (the gate g(f, ρ), the AM-style telic-absorption term, the exponential ρ(t) trajectory) and a piecewise-linear placeholder where it does not (relational dose-response above d_safe).
- **Integration.** The three lit-review domains (work, relationships, meaning) and the three reframer mechanisms from the topology (G3, G4, G7) are unified under a single equation rather than treated as parallel stories. The integration is what gives the strategic recommendations their leverage: S3 ballast moves B, S5 dose-limit moves d, S6 effortful practice moves ρ — and because ρ is shared across channels, S6 is structurally the highest-leverage move even though it looks like a workplace-productivity tip.

What is **ready for formalisation**:

| Piece | Source | Form |
|---|---|---|
| Productivity scale α | Brynjolfsson-Li-Raymond, BCG RCT | α ≈ 0.40, gated by f·ρ |
| Self-automator gate τ | Randazzo BCG self-automator share + skill-development null | τ ≈ 0.30, σ ≈ 0.06 |
| Telic-absorption ΔM (G7) | Setiya telic/atelic + topology A5 | −κ·φ·max(0, T−B) |
| Competence-erosion ΔM (G4) | SDT competence-frustration + topology G10 | −λ_M·(1−ρ)·T |
| Relational dose-response | OpenAI-MIT N=981 | piecewise: protective ≤ d_safe, harm > d_safe |
| Time evolution of ρ | A3 crux made parametric | ρ(t) = ρ₀·exp(−λ·u·t) |

Two of the topology's reframer mechanisms (G4 cognitive offloading bridge, G7 telic exhaustion) admit mechanistic closed-form representations as listed above. The third reframer (G3 engagement-optimized substitution) is *not* represented mechanistically — it is a *parameter source* for ψ_R: G3 informs the calibration of the dose-response slope above d_safe, but the model has no machinery for the engagement-optimization mechanism itself. If G3 weakens (regulation, consumer demand, competitor design), ψ_R would re-calibrate but the model's structure would not change. This is a real scope difference: G4 / G7 are encoded in the model; G3 is encoded only in the calibration of one parameter.

What is **still observational** and inherited by Stage 4 as fitting targets:

- λ (cumulative-atrophy speed) — the single largest unknown. The topology's open question O3 is exactly "is λ ≈ 0 or λ ≈ 0.06?" The 2+ year longitudinal study that would answer this does not yet exist; Stage 4 will use the cross-sectional evidence (Gerlich, Stadler-Bannert-Sailer, Kosmyna, Ehsan) to bound λ from below but cannot pin it down.
- ψ_R (above-d_safe slope), β_R (below-d_safe slope), and the precise location of d_safe — the OpenAI-MIT data give the qualitative shape but the magnitudes will need refitting from the 300,000-message dataset directly.
- κ (competence-frustration sensitivity) calibration across populations and trait classes. The SDT literature gives within-study coefficients but not a portable scale.
- The scalar-identity assumption on T and B. The topology treats meaning architecture as one thing; in reality identity is multi-domain (work, family, hobbies, civic), and the AI absorption fraction φ may differ across components. Stage 4 should test whether scalar T is a useful approximation or whether a vector T = (T_work, T_family, T_hobby, …) is required.

**Time horizon.** ΔV and ΔM are comparative-statics changes — they are not instantaneous nor lifetime-cumulative. The defensible reading is that the static dashboard's ΔNet represents change over a "medium" horizon, ~1–5 years, where the fast channels (ΔV_prod from current AI capability) and slow channels (ΔM_telic, ΔM_comp from prolonged absorption and offloading) meaningfully overlap. Shorter than ~1 year, the productivity-gain side over-weights because the meaning-erosion channels haven't had time to materialise. Longer than ~5 years, the trajectory tab — which evolves ρ explicitly — is the better lens because the static parameters drift. The horizon is implicit because the literature does not yet give a single time-scale for the slow channels (the cross-sectional evidence is duration-vague). Stage 4 should test whether different empirical anchors imply different horizons and, if so, build a time-resolved version of the dashboard rather than a static one.

## 2. The generating function

### 2.1 Spine

For an individual, over a fixed horizon Δt:

```
ΔNet = ΔV + ΔM
ΔM ≤ 0 by construction
```

ΔV is offensive — the value gained from AI augmenting projects, accessing previously-out-of-reach domains, and (in the relational channel) the therapeutic-grade benefit at low dose. ΔM is defensive — the meaning lost when AI absorbs work that was carrying meaning, when retained practice atrophies and competence erodes, and when high-dose AI-emotional engagement substitutes for in-person relationships in a thin baseline.

**A note on the cut.** The V/M split is an organising convention for presentation, not a structural claim about the model. The substantive structure is the six-channel additive decomposition; "ΔV" and "ΔM" group channels by typical sign for ease of reading. ΔV_trap in particular is bounded above by 0 (it is always negative-or-zero) and could equivalently be moved to the ΔM side without changing any of the model's claims. The block-coupling through ρ runs across the V/M cut: ΔM_comp (defensive side) and ΔV_prod / ΔV_trap (offensive side) all move together when ρ moves. If you find the V/M framing more confusing than helpful, read the channels directly and ignore the cut.

The spine is intentionally minimal. The work happens in the channel-level decomposition.

### 2.2 The three defensive channels (ΔM)

```
ΔM_telic = −κ · φ · max(0, T − B)
ΔM_comp  = −λ_M · (1 − ρ) · T
ΔM_rel   = −ψ_R · max(0, d − d_safe) · (1 − δ_R)
ΔM       = ΔM_telic + ΔM_comp + ΔM_rel
```

**ΔM_telic — telic absorption.** The Setiya / topology A5 channel. Fraction φ of telic work that AI absorbs, weighted by competence-frustration sensitivity κ, conditional on the telic share T exceeding the atelic ballast B. When B ≥ T, the meaning that was staked on telic completion has somewhere else to live; when T ≫ B, telic absorption is structurally costly. The max(0, ·) is the topology handoff's "B ≥ T → ΔM ≈ 0" boundary made explicit.

**ΔM_comp — competence erosion (the bridge).** The G4 / G10 channel. Practice atrophy degrades competence not just symbolically (AI can do it) but actually (you cannot do it any more). The (1 − ρ) factor is the inverse of retained practice; the T multiplier is because competence erosion only hurts to the extent the lost competence carried meaning. This is the channel that makes ρ the bridge: a single parameter pulls both the offensive gate (below) and the defensive ΔM_comp term.

**ΔM_rel — relational dose-response.** The G3 / E11 / E12 channel. The OpenAI-MIT N=981 RCT found loneliness, dependence, and reduced in-person socialization at high daily voluntary use; below a threshold (placed at d_safe in the model), the same engagement is protective rather than harmful. Above d_safe, the penalty scales linearly in the excess dose, dampened by the relational baseline thickness δ_R — someone embedded in thick local relational infrastructure absorbs the dose better than someone in the Anti-Social-Century baseline. ψ_R is the per-minute slope; in the dashboard it is calibrated so that 60 minutes above d_safe in a thin baseline produces a meaningful but not catastrophic ΔM_rel.

### 2.3 The three offensive channels (ΔV)

```
g(f, ρ) = 1 / (1 + exp(−(f · ρ − τ) / σ))

ΔV_prod = g · α · a · (1 − s)
ΔV_rel  = β_R · min(d, d_safe) · (1 − δ_R / 2)
ΔV_trap = −(1 − g) · η_trap · a
ΔV      = ΔV_prod + ΔV_rel + ΔV_trap
```

**g(f, ρ) — the self-automator gate.** The smooth analogue of the topology handoff's "above thresholds → upside, below → trap" dichotomy. The product f · ρ is the relevant axis: feedback-loop richness times retained practice. Both have to be present for AI use to upskill; if either collapses, the gate closes. τ ≈ 0.40 puts the threshold roughly where Randazzo's self-automator boundary lies in the BCG data; σ ≈ 0.06 makes the transition smooth enough that small parameter changes don't flip the regime discontinuously.

**ΔV_prod — productivity / novice-skill compression.** Brynjolfsson-Li-Raymond's customer-service finding (+14% overall, +34% novice) plus the BCG RCT inside-frontier gains. The (1 − s) factor is novice-skill compression: experts gain little from AI on tasks they're already good at, novices gain the most. α is the productivity scale; in the dashboard, α = 0.40 reproduces Brynjolfsson novice gain at s ≈ 0.2 (α · 0.8 ≈ 0.32, in the 30–34% range). Multiplied by a (AI capability on the task) — when capability is low or task is outside the jagged frontier, ΔV_prod shrinks regardless of s.

**ΔV_rel — therapeutic-grade relational benefit.** Therabot RCT (Heinz et al. 2025) and the De Freitas equivalence-hypothesis cluster — at low dose, AI emotional engagement produces real benefit. Capped at d_safe so the channel does not double-count the dose-response harm; the (1 − δ_R/2) factor flattens the benefit for people in already-thick relational baselines (they get less marginal benefit from low-dose AI use because they have alternatives).

**ΔV_trap — self-automator penalty.** When the gate closes (g → 0), AI use is not just unhelpful but actively degrading: the user delegates without the practice and feedback loops that would catch errors, and incurs both the productivity gain they fail to capture and the deskilling cost. η_trap is the penalty scale; calibrated so that f = ρ = 0.2 (well below threshold) and a = 0.8 produces a moderately negative ΔV_prod + ΔV_trap sum even before adding ΔM.

### 2.4 The bridge through ρ

The cross-domain bridge G4 from the topology is encoded as one fact: ρ enters three of the six channels — one defensive directly, two offensive via the gate.

```
ΔM_comp = −λ_M · (1 − ρ) · T          ← direct: ρ low → competence erosion
g       = sigmoid((f · ρ − τ) / σ)    ← intermediate: ρ low → gate closes
ΔV_prod = g · α · a · (1 − s)         ← gated: lower g → less productivity gain
ΔV_trap = −(1 − g) · η_trap · a       ← gated: lower g → larger trap penalty
```

The remaining three channels (ΔM_telic, ΔM_rel, ΔV_rel) do not depend on ρ. ΔM_telic depends on (T, B, φ, κ) — the identity-allocation and AI-exposure parameters. ΔM_rel and ΔV_rel depend on (d, δ_R) — the daily-relational-dose and baseline-thickness parameters. Lowering ρ has compounding effects across the three ρ-coupled channels: it directly increases ΔM_comp, and it lowers g which simultaneously reduces ΔV_prod and increases ΔV_trap. This is the mathematical content of the topology's claim that S6 (maintain effortful practice) is load-bearing — it is a single-parameter intervention with leverage in three of the six channels.

The bridge is also why the channel-level decomposition is *not* fully orthogonal even though each channel is written as an additive term. The three ρ-coupled channels are **block-coupled through ρ**: at fixed ρ, they are independent; varying ρ moves them in coordinated ways. This is the model's representation of the lit review's structural finding that the work-deskilling and meaning-competence channels share a generating function (G4) — they are not separate transitions but one transition with two projections. The topology's broader claim that *all three* domains (work, relationships, meaning) share a generating function is a stronger version that the model represents weakly: ρ couples the work and meaning channels directly, while the relational channel is coupled only via the topology's parallel-mechanism inference (which the model leaves unencoded; see Objection 2 in §7 for why).

### 2.5 Time evolution

The trajectory tab makes A3 (cumulative offloading) parametric:

```
ρ(t) = ρ₀ · exp(−λ · u · t)
ΔNet(t) = compute with ρ(t); all other parameters held fixed
```

Where λ is the atrophy speed and u is the offloading rate. λ = 0 is the calculator-analogue: using AI is structurally like using a calculator and retained practice does not decay. λ > 0 is the cumulative-atrophy regime: ρ falls exponentially in the offloading-rate × time product. Only ρ evolves; pre-attempt skill stock s, AI capability a, feedback richness f, and the identity / relational parameters are held fixed across the trajectory. The deskilling-over-time effect is captured *through* ρ, not as a separate s(t) decay: as ρ drops, the gate g(f, ρ) closes, which propagates to ΔV_prod (smaller productivity gain) and ΔV_trap (larger penalty), while ΔM_comp grows directly. Adding a second decaying variable would either double-count the deskilling or imply the user becomes a fresh novice with full upside available — neither of which matches the lit-review evidence.

The empirical question O3 from the topology is which regime is correct for AI-augmented knowledge work. The dashboard makes the answer visibly consequential: under λ = 0 (calculator analogue), a knowledge worker with ρ₀ = 0.8 has flat ΔNet over 10 years; under λ = 0.06 (heavy cumulative regime, ρ-half-life ≈ 19 years), the same person has ρ(10) ≈ 0.56, the gate closes from g ≈ 0.84 to g ≈ 0.42, and ΔNet drifts measurably more negative as ΔV_prod compresses and ΔV_trap grows — the drift magnitude over 10 years is on the order of −0.15 ΔNet under default identity parameters (T = 0.7, B = 0.2). Meaningful, but not catastrophic; cumulative-atrophy is a slow process even in its strong regime, and the dramatic-collapse intuition some readers bring to "AI rots your brain" framings does not map onto the model's actual trajectory at calibrated λ. The model cannot answer which regime is right; it can only make the difference legible.

## 3. Closed-form pieces

Three components admit clean equations. The rest are calibrated approximations.

### 3.1 The gate g(f, ρ)

```
g(f, ρ) = 1 / (1 + exp(−(f · ρ − τ) / σ))
```

Standard logistic with location τ and scale σ. The product f · ρ on the location axis is the substantive choice: feedback-loop richness alone is not enough (you can have great feedback loops but no practice if you've fully delegated to AI), and retained practice alone is not enough (you can grind without ever testing your work against reality). Both must be present.

Worked anchors at τ = 0.30:
- f = 0.7, ρ = 0.7 (cyborg / centaur centroid in BCG): f·ρ = 0.49 → g ≈ 0.96. Strong upside regime.
- f = 0.5, ρ = 0.5 (median knowledge worker): f·ρ = 0.25 → g ≈ 0.30. Just below the inflection — small interventions push them above; small drift pushes them below.
- f = 0.2, ρ = 0.3 (full self-automator): f·ρ = 0.06 → g ≈ 0.02. Trap effectively closed.

The threshold τ = 0.30 is a structural-claim choice, not a tight fit. The substantive claim is "the median knowledge worker without conscious intervention is *on the edge* — small movements in f or ρ push them either way" — which the topology supports but no single dataset cleanly calibrates. The BCG-Randazzo 27% self-automator share gives a rough lower-bound anchor (someone with f ≈ ρ ≈ 0.2–0.3 should be in the trap), and the centaur upper-tail (f ≈ ρ ≈ 0.7–0.8) gives a rough upper-bound anchor (clearly above the gate) — but the precise location of τ between these points is a modelling choice. Pass 1 used τ = 0.40 and produced g ≈ 0.07 at the median, which encoded a stronger structural claim than the data supports ("most knowledge workers are deeply trapped"); pass 2 lowered to τ = 0.30 to encode the weaker, better-supported claim ("the median is on the edge"). Stage 4 should test τ directly against the BCG individual-level data; whether the value generalises beyond consulting work is a separate question that requires non-BCG samples to answer. σ = 0.06 makes the transition smooth (10–90% transition over an f·ρ range of about 0.27, or roughly two standard deviations of within-population variation in either parameter alone).

### 3.2 Telic-absorption ΔM_telic

```
ΔM_telic = −κ · φ · max(0, T − B)
```

Three multiplicative factors. **φ ∈ [0, 1]** is the AI-absorbable fraction of telic work — for software engineering φ might be 0.6 (much of the implementation absorbed but design judgment retained); for in-person therapy φ might be 0.1. **(T − B) ∈ [−1, 1]** is the unballasted telic share — how much of identity is staked on telic completion that does not have an atelic counterpart. **max(0, ·)** truncates the negative branch: if B exceeds T, atelic ballast covers the whole telic identity-share and then some, so absorbing telic work does not pull meaning out of unfilled space. **κ ∈ [0, 1]** is competence-frustration sensitivity — the SDT-derived scalar that translates competence shortfall into amotivation. Some people are temperamentally more sensitive to this; the model exposes κ as a slider rather than a constant.

The boundary B = T is the topology's atelic-ballast hypothesis (S3) made structurally precise. It is **not** that ballast eliminates the productivity threat; it is that ballast moves the meaning-architecture out of the line of fire. ΔV_prod is unchanged by raising B; only ΔM_telic shrinks. This is what S3 actually buys.

### 3.3 Competence erosion ΔM_comp (the bridge term)

```
ΔM_comp = −λ_M · (1 − ρ) · T
```

The (1 − ρ) factor is offloading: the fraction of capacity that has slipped because practice was delegated. The T factor is because competence erosion only matters to the extent the eroded competence was carrying identity. λ_M is the coefficient — calibrated so that ρ = 0, T = 1 produces ΔM_comp = −0.30 (a severe SDT-style competence-frustration hit, the upper end of what BPNSFS scales operationalise).

Note the asymmetry with ΔM_telic. Building atelic ballast (raising B) is **no defense against ΔM_comp** — the competence erosion is happening in the telic domain regardless of where else identity is anchored. The only defense against ΔM_comp is keeping ρ high (S6 effortful practice), which is exactly the topology's load-bearing recommendation. The model makes this asymmetry visible: pulling B up zeroes out ΔM_telic but leaves ΔM_comp untouched; pulling ρ up shrinks both.

### 3.4 Relational dose-response

```
ΔM_rel = −ψ_R · max(0, d − d_safe) · (1 − δ_R)
ΔV_rel = +β_R · min(d, d_safe) · (1 − δ_R / 2)
```

Piecewise-linear, with the kink at d_safe = 30 minutes/day. Below d_safe, the dose carries therapeutic-grade benefit (Therabot scale, scaled by β_R = 0.001). Above d_safe, every additional minute carries a dose-response penalty (ψ_R = 0.003 per minute per (1 − δ_R)). At the maximum below-threshold dose (d = d_safe = 30) in a fully thin baseline (δ_R = 0), peak benefit is β_R · 30 · 1 = 0.030 — comparable to ΔV_prod at moderate productivity gain. Pass 1 used β_R = 0.004 which made peak benefit ≈ 0.12, dominating ΔV_prod even in well-functioning AI-use scenarios; that overstated the generalisation from Therabot's clinical-population effect to the median user.

Two things this form does not capture and which Stage 4 should refit:
- **Possible sigmoidal saturation at very high doses.** ψ_R may fall once d gets large enough — at some point further engagement does not produce *additional* harm because the relational substitution is already complete. The OpenAI-MIT data is too narrow in dose range to show this.
- **The crossover may be nonlinear.** d_safe is a sharp kink in the model; in the data it is plausibly a smooth U-shape. The kink is a useful caricature for the dashboard but is not a structural claim about the curve.

The (1 − δ_R) and (1 − δ_R/2) factors encode the topology's A4 (relational depletion is structural) crux: in a thin relational baseline, both the upside and (especially) the downside are amplified. In a thick baseline, both attenuate. The asymmetric attenuation (penalty fully scaled by δ_R, benefit only half-scaled) is the model's claim that thick baselines mostly *protect* — they don't crowd out low-dose benefit but they do dampen high-dose harm.

### 3.5 Productivity gain ΔV_prod

```
ΔV_prod = g(f, ρ) · α · a · (1 − s)
```

Already covered above. Worth flagging that the (1 − s) novice-skill-compression factor is the offensive-side analogue of the (T − B) ballast structure on the defensive side: just as ballast determines whether telic absorption is costly, pre-attempt skill stock determines whether AI augmentation is large or marginal. **The asymmetric-exploiter scenario in the dashboard is exactly the configuration where (1 − s) is large, f and ρ are above the gate, and the upside dominates** — a novice with the discipline to maintain feedback loops and effortful practice is structurally better-positioned than an expert who delegates without either.

**Cross-domain α heterogeneity.** The model's α = 0.40 is a midpoint compromise across studies that span ~6× in measured productivity gain (α ≈ 0.24 to 1.4). Translating empirical "+X% productivity" into α uses ΔV_prod = α·a·(1−s) at gate-open: α = ΔV_prod / (1−s) with assumed average s for the study population.

- **Brynjolfsson-Li-Raymond customer service** is a two-anchor study. +14% overall productivity at average s≈0.5 → α ≈ **0.28**. +34% novice (s≈0.2) → α ≈ **0.43**. The within-study spread is 1.5× — already non-trivial — and the model's α=0.40 picks up the novice-end of this range.
- **BCG consulting** (Dell'Acqua-McFowland-Mollick) is two-outcome: +12% productivity inside the frontier (s≈0.5) → α ≈ **0.24**; +40% quality on the same tasks → α ≈ **0.80**. Productivity-vs-quality spread within a single study is ~3×, the largest within-study spread on record.
- **Cui-Demirer-Jaffe-Musolff-Peng-Salz** Copilot/coding field experiments found ~26% increase in completed-task throughput per week. With developer s ≈ 0.4–0.5: α ≈ **0.5–0.65**.
- **Peng-Kalliamvakou-Cihon-Demirer** GitHub Copilot controlled experiment found 55% faster task completion on a JavaScript HTTP-server task. With s ≈ 0.4–0.5: α ≈ **1.0–1.4**.

The honest reading: α spans ~0.24 to ~1.4 across measured studies — a 5–6× range, not the 4× I previously claimed. The model's α = 0.40 picks up roughly the median of this range and is closest to the Brynjolfsson-novice anchor. Users in highly AI-suited domains (coding, writing, image generation, data analysis) should mentally bump α up by 1.5–3×; users in BCG-productivity-class tasks should bump it down by ~1.5×; users in domains where AI is awkward (in-person therapy, embodied physical work, deep relational work) should bump it down further still. Stage 4 should fit α per major task category rather than treating it as a scalar — see Q6 in §9.

### 3.6 ρ(t) trajectory

```
ρ(t) = ρ₀ · exp(−λ · u · t)
```

Standard exponential decay. λ is the per-unit-offloading atrophy rate; u is the offloading rate. Their product is the effective atrophy speed. Half-life = ln(2) / (λ · u) — under λ = 0.06 and u = 0.6 (heavy offloading, cumulative regime), half-life ≈ 19 years; under λ = 0.04, half-life ≈ 29 years; under λ = 0.02, half-life ≈ 58 years. Even in the cumulative-atrophy regime, the timescale is multi-decadal, which is consistent with intuition rust being a slow process.

Only ρ evolves in the trajectory tab; pre-attempt skill stock s, AI capability a, feedback richness f, and the identity / relational parameters are held fixed across the horizon. This is a deliberate simplification (pass 1 had s also decaying, but the asymmetric exponential-vs-linear decay was unprincipled, and adding a second decaying variable doubles-counts deskilling-over-time which is already fully captured through ρ → g closing → ΔV_prod giving way to ΔV_trap).

The form is phenomenological — λ is not derived from a deeper biological model. It's a calibration knob whose value is the falsification window for A3. Stage 4 will bound it from below using the existing cross-sectional evidence; pinning it down requires longitudinal data that does not exist yet.

## 4. Composing the parts: scenario anchors

The dashboard's six presets are calibration targets — each should produce a recognisable position from the lit review.

| Preset | T | B | φ | κ | s | a | f | ρ | d | δ_R | Expected output (τ=0.30, β_R=0.001) |
|---|---|---|---|---|---|---|---|---|---|---|---|
| Default risk | 0.70 | 0.20 | 0.60 | 0.50 | 0.60 | 0.70 | 0.50 | 0.60 | 15 | 0.40 | g ≈ 0.50 (on the threshold); ΔV slightly negative (trap and productivity roughly cancel); ΔM dominated by telic absorption + competence erosion; ΔNet ≈ −0.27 |
| Ballast intervention | 0.50 | 0.50 | 0.60 | 0.50 | 0.60 | 0.70 | 0.50 | 0.60 | 15 | 0.40 | Same g, same ΔV; ΔM_telic → 0 (B = T); ΔM_comp unchanged; ΔNet ≈ −0.10 (≈ +0.17 improvement over default risk) |
| Self-automator | 0.60 | 0.30 | 0.70 | 0.40 | 0.40 | 0.80 | 0.20 | 0.30 | 10 | 0.40 | g ≈ 0.02 (gate effectively closed); ΔV_prod ≈ 0; ΔV_trap ≈ −0.24; ΔNet ≈ −0.43 |
| Asymmetric exploiter | 0.40 | 0.40 | 0.40 | 0.40 | 0.20 | 0.80 | 0.70 | 0.70 | 5 | 0.50 | g ≈ 0.96; large ΔV_prod from (1−s = 0.8); ΔM_telic = 0; modest ΔM_comp; ΔNet ≈ +0.20 |
| Creative pre-AI | 0.80 | 0.10 | 0.80 | 0.70 | 0.70 | 0.80 | 0.40 | 0.50 | 10 | 0.30 | g ≈ 0.16; small ΔV (high s + partial gate, mostly trap penalty); very large ΔM_telic ≈ −0.39 from κ·φ·(T−B) compounding at extremes; displacement-anxiety trajectory; ΔNet ≈ −0.69 |
| Heavy companion | 0.50 | 0.30 | 0.40 | 0.40 | 0.50 | 0.70 | 0.50 | 0.60 | 90 | 0.20 | g ≈ 0.50; relational channel dominant — ΔM_rel ≈ −0.14 from 60-min dose excess in thin baseline — but s = 0.50 gives more (1−s) upside than the default-risk preset, partially offsetting ΔV_trap; ΔNet ≈ −0.24 |

Three non-obvious calibration targets:

1. **The default-risk preset should produce ΔNet < 0 driven by the meaning side, not the gate.** Under the new τ = 0.30 the gate sits at g ≈ 0.50 for default risk — ΔV is approximately neutral, and the negative ΔNet comes from ΔM_telic + ΔM_comp. This is the right structural reading: the default knowledge-worker risk is not that AI use is broken (the gate isn't fully closed) but that meaning is leaking from the unballasted telic identity and from competence erosion at moderate ρ.
2. **Ballast intervention should improve ΔNet without changing ΔV.** Raising B alone zeroes ΔM_telic; nothing else changes. The dashboard makes this visible: the ΔV bar is identical between default risk and ballast intervention, only the ΔM_telic channel collapses.
3. **The asymmetric-exploiter preset should beat the ballast intervention by a large margin.** ΔNet ≈ +0.20 vs ≈ −0.10 — the offensive path is not just "different" from the defensive path, it dominates when the conditions are met. The point of the upside framing is to make this comparison legible; a user comparing the two presets should see that S3 (ballast) is a partial mitigation while the asymmetric-exploiter regime is a positive-sum strategy.

## 5. Boundary conditions and where the model breaks

Five boundaries are explicit:

1. **B ≥ T (atelic ballast covers telic identity).** ΔM_telic = 0 by construction. The model still computes ΔM_comp and ΔM_rel; ballast does not protect against competence erosion or relational dose-response. This is the topology's S3 strengthening exactly the channel S3 was designed to address.

2. **f · ρ ≪ τ (gate fully closed).** g ≈ 0; ΔV_prod ≈ 0; ΔV_trap ≈ −η_trap · a. Below the gate, AI use is a net negative on the offensive side regardless of capability a. The model is making the strong claim that there is a regime where more AI capability *increases* damage rather than reducing it. This is not a bug; it is the formalisation of the self-automator finding.

3. **s → 1 (expert at the relevant task).** ΔV_prod → 0. AI augmentation offers little to someone already fully capable. The model is silent on the *meaning* dimension here — an expert may still suffer ΔM_telic and ΔM_comp from AI absorbing their craft even when they get no productivity benefit from it. This is a real and well-attested experience (creative-displacement anxiety in expert practitioners) and the model captures it correctly: ΔV ≈ 0 while ΔM is large and negative.

4. **a → 0 (AI cannot do this task).** Both ΔV_prod and ΔM_telic shrink — there is no absorption when AI cannot do the work. The relational channels are independent of a (they depend on d). This boundary is what the topology calls the "outside the jagged frontier" region.

5. **λ = 0 (calculator-analogue).** ρ(t) = ρ₀ for all t; the trajectory tab collapses to a flat line. This is one tail of the A3 crux. The model under this regime is still useful — the cross-sectional ΔNet is still meaningful — but the time evolution becomes trivial and S6 (effortful practice) loses its strong defensive role (it still maintains the gate and ΔM_comp at t = 0, but does not protect against compounding). The dashboard's "calculator" preset on the trajectory tab makes this regime explicitly testable.

Beyond these, five structural limits are not boundary conditions but scope limits:

- **Individual scope only.** The model takes T, B, φ, etc. as individual parameters. Aggregate dynamics (labor-share macro, civilisational meaning architecture) are out of frame, matching the topology's individual-decision-maker scope.
- **Western-population calibration.** All anchor values come from US/Western studies (BCG, Brynjolfsson, OpenAI-MIT, Therabot, Anti-Social Century, Common Sense). The model's predictions for non-Western contexts should be treated as hypotheses, not findings — particularly for δ_R (relational baseline) which is most sensitive to cultural variation.
- **Single-task framing of φ and a.** The model treats AI capability and AI absorbability as scalars over the user's telic work. Real careers span multiple tasks with very different φ and a. A vector form (φ_i, a_i) is a future extension; the scalar version is sufficient for the structural claims.
- **Labor-market access is exogenous (G5 apprenticeship-break).** The model takes "you have access to the work" as given and computes ΔV_prod from there. But one of the strongest empirically-supported labor findings in the lit review is the apprenticeship-ladder break — entry-level employment in highly AI-exposed occupations fell ~13% (Brynjolfsson-Chandar-Chen ADP data, Aug 2025) for 22-25-year-olds, while same-occupation employment for over-35s rose. For early-career users in highly-exposed occupations, *getting the work in the first place* is the dominant practical concern, and the model is silent on it. ΔV_prod estimates for novices should be read as conditional on labor-market access; if you don't have the access, the model's offensive side does not apply and S7 (career bet) is not the right frame to begin with. This is a structural feature of the labor market, not a parameter the individual can pull.
- **Material-floor primacy (L2).** For users whose material floor is insecure (precarious income, dependent-care obligations, no economic runway), the absence of ΔV_prod — i.e., job loss or income disruption — is the existential question. The model's individual-decision framing implicitly assumes economic security: ΔM_telic versus ΔV_prod is a choice you can have only when survival is not in question. The topology's logical guardrail L2 names this directly, and the model inherits it as a scope limit. For users at the material-secure end of the distribution, the model's strategic recommendations are decision-relevant; for users at the material-insecure end, the labor-economics question dominates and the meaning-architecture machinery here is not the right tool.

## 6. Distortion-aware reading

Each channel of the decomposition has a public-discourse failure mode. The model's job is to make the failure visible.

| Channel | Common misreading | What the model says |
|---|---|---|
| ΔV_prod (positive) | "AI is making everyone more productive" | Conditional on g(f, ρ) being open. Below the gate, the same a and s give negative ΔV via ΔV_trap. Productivity gains are real *but only for users meeting feedback / practice conditions*. |
| ΔM_telic (negative) | "AI is destroying meaning" | Conditional on T > B. The atelic-ballast intervention (S3) zeroes this channel without changing AI exposure. The threat is identity-allocation-dependent, not AI-dependent. |
| ΔM_comp (negative) | "AI use rots your brain" | Conditional on (1 − ρ) being large *and* T being large. People who maintain practice (high ρ) take little hit even at high AI exposure; people whose identity isn't telic-loaded (low T) take little hit even at low ρ. The bridge is real; its blanket framing is wrong. |
| ΔM_rel (negative at high d) | "Companion apps are bad" | Conditional on d > d_safe and δ_R low. Below d_safe, the same engagement is therapeutic-grade benefit (Therabot). The harm is dose-and-baseline-dependent, not modality-dependent. |
| ΔV_rel (positive at low d) | "AI is great for mental health" | Conditional on staying below d_safe. The same engagement above d_safe flips sign. The dose-response framing is the only honest summary. |
| ΔV_trap (negative when below gate) | "Some people just don't know how to use AI" | The trap is structural, not cultural. f and ρ are observable conditions; users meeting them upskill, users not meeting them deskill, regardless of intent or attitude. |

The four topology distortions (D1 fatalism, D2 slow-camp dismissal, D3 productivity-only optimization, D4 material-blind class bias) each map to a specific failure mode of reading the decomposition: D1 ignores ΔV (assumes the gate closes globally); D2 ignores ΔV_prod and ΔM_comp (claims aggregate effects haven't moved); D3 ignores ΔM and ΔV_trap (optimises only ΔV_prod); D4 ignores L2 / the model's silence on material floor.

## 7. Adversarial + steelman

Five objections to the formalisation itself.

### Objection 1 — Variance bookkeeping for meaning is a category error

ΔM is measured on the same axis as ΔV, but meaning is not a quantitative quantity that admits subtraction or comparison across persons. Calling something "−0.30 ΔM" is pseudo-precision: meaning is qualitative, embodied, narratively structured. The decomposition smuggles in a positivism that the underlying phenomenon does not support.

**Steelman.** This is the strongest version of the philosophical-versus-quantitative tension. The lit review's meaning section deliberately uses Setiya, Arendt, and SDT — three traditions, none of which treat meaning as scalar. Reducing them to one number is exactly what those traditions warn against. The objection has force precisely because the model is *not* claiming to measure meaning literally; it is claiming that comparative claims about meaning ("this configuration is better-positioned than that one") admit a structured representation, and the structure is the model's content.

**Response.** Conceded in part. The model is more accurately described as a **comparative-meaning generating function**: given two parameter configurations, it produces a directional claim about which is structurally better-positioned. The cardinal numbers ΔM = −0.20 vs −0.50 should be read ordinally, not metrically. Anyone who reads the dashboard as "you will lose 20% of your meaning" is misreading; the dashboard's job is to surface the channels and their interactions, and the bare numbers are pedagogy for that job. Stage 4 will calibrate the channels against measurable proxies (SDT subscale shifts under intervention, time-use-survey-derived telic/atelic share changes, OpenAI-MIT loneliness-scale movements) without claiming the result is meaning-measurement.

### Objection 2 — Encoding the bridge as a single ρ over-collapses three different mechanisms

Cognitive offloading in analytical work, in emotional processing, and in identity-supporting practice are three different psychological processes with different timescales and different reversibilities. Treating them as one parameter ρ is the topology's bridge made artificially clean — the bridge is *an inference*, not an observed equivalence, and the cross-sectional evidence (Gerlich, Stadler-Bannert-Sailer) is for the analytical channel only.

**Steelman.** This is correct as stated — the topology itself flags that channels 2 (relational) and 3 (meaning) are parallel-mechanism inferences, not directly observed. Encoding them as one parameter is a load-bearing simplification, not a theorem. The simplification has structural consequences: pulling ρ in the dashboard moves all three channels in lockstep, which over-states the actual coupling. A model with three ρs (ρ_analytic, ρ_emotional, ρ_meaning) would let the channels move independently and would more accurately represent the evidence.

**Response.** Conceded with a scope limit. The single-ρ form is the appropriate first formalisation: the topology argues the three channels share a generating function (G4), and a one-parameter representation is the minimum-complexity form consistent with that argument. The dashboard's bridge slider is meant to test the strong form of the argument. Stage 4 should test whether ρ can be split — if O2 (asymmetric-adoption couples) shows that emotional ρ moves independently of analytical ρ, the model needs a vector form. The single-ρ version is falsifiable in exactly that way, which is part of what makes it useful.

### Objection 3 — Linear ΔM_rel above d_safe is not what the OpenAI-MIT curve looks like

OpenAI-MIT N=981 reports a continuous monotone-increasing relationship between daily voluntary use and loneliness, dependence, and reduced in-person socialisation. There is no clean kink at "30 minutes/day"; the safe-threshold is a model artifact. Plus the literature on dose-response in companion-app harm (de Freitas identity-discontinuity) suggests *non-linear* damage at high doses (the worst harm is from the loss of access, not the use itself), which the model's piecewise-linear form fully misses.

**Steelman.** Both points are correct. The kink at d_safe is a useful pedagogical device — it makes the dose-response visible and binds it to a recognisable threshold — but the underlying curve is plausibly smooth and almost certainly nonlinear at the high tail. Heavy users may have already substituted away from human relationships, so additional minutes do not produce additional substitution; this would mean ψ_R falls at high d. Conversely, the *catastrophic-loss* mechanism (Replika-removal evidence) is not a function of d at all — it is a function of platform stability — and the model has no representation of this.

**Response.** Conceded fully on both points. The piecewise-linear form is a placeholder; Stage 4 should fit the curve directly from the OpenAI-MIT messaging dataset and the de Freitas survey panels. The catastrophic-loss mechanism is currently outside the model's scope — it would need a separate "platform-shock" term that doesn't fit the additive-channel structure cleanly. For now, the model captures the steady-state dose-response but not the failure-mode dynamics; a Stage-3 extension or a Stage-5 build artifact may add this.

### Objection 4 — Treating identity as scalar T and B misses the multi-domain structure

Real people allocate identity across work, family, civic, hobby, and friendship domains. Each domain has its own telic/atelic balance and its own AI absorption fraction. Compressing all of this to scalar (T, B) is the modelling equivalent of treating a person as having "one personality trait." The *whole point* of the topology's S2 (identity diversification) recommendation is that diversification across domains is the move; a scalar model cannot represent that move.

**Steelman.** This is correct. The dashboard's "raise B" lever is doing two different things at once: it's literally increasing atelic share, but it's also implicitly *diversifying* identity (because increasing atelic share is the same operation as adding a non-telic identity domain). The model conflates the two and the conflation is structural — the additive ΔM_telic = −κ·φ·max(0, T−B) form does not distinguish "more domains" from "more atelic share within the same domain." S2 is therefore representable in the dashboard only if you read raising B as proxy for raising domain count.

**Response.** Conceded. A vector form **T** = (T_work, T_family, T_civic, T_hobby, …) with corresponding **B**, **φ**, **a**, **κ** is the right next step. The scalar version is the appropriate first cut: it captures the structural argument (telic absorption hurts, atelic ballast helps, the bridge runs through ρ) cleanly, and the four objections do not break the structural argument — they refine its precision. Stage 4 should test whether the scalar approximation produces reasonable agreement with empirical configurations of real careers, and Stage 5 may build a vector-form interface if the scalar approximation is too lossy.

### Objection 5 — The model is moralistic productivity discourse in equations

Look at what the model rewards: high f (rich feedback loops), high ρ (maintained effortful practice), high B (built atelic ballast), low d (limited AI-emotional dose). Every parameter that the model says "moves you toward better outcomes" is a parameter that mainstream AI productivity culture already moralises about. The disciplined user wins on every channel. This is not structural analysis of the AI transition; it is a sophisticated rephrasing of "be diligent, build hobbies, don't over-rely on AI" — advice you could find in any 2025 LinkedIn post, dressed up in Greek letters and a sigmoid gate. The model's apparent mathematical content is, on this reading, a credentialing operation: it gives the prevailing productivity-culture prescriptions an air of formalism without changing what they actually recommend.

**Steelman.** The objection is sharper than D3 (productivity-only optimization, which targets readers who *ignore* the relational and meaning channels). D3 says "don't optimise only for ΔV_prod"; this objection says "even when you include all six channels, the model rewards conscientiousness uniformly, and that uniform reward is the moral content smuggled into the structure." The model has no parameter for "your domain is being eaten regardless of your effort." It has no representation for "personal discipline cannot save you from labor-market disruption." The G5 scope limit gestures at this in §5 but the dashboard itself has no surface where structural-disruption-overwhelms-personal-effort can be made visible. A reader pulling sliders sees "discipline → better outcome" everywhere they look, which is exactly the conclusion the productivity-culture frame wants them to reach. Independently, the Caporusso "creative displacement anxiety" finding documents practitioners who suffer despite high discipline, high feedback richness, and intact relational baselines — the dashboard's creative-pre-AI preset captures this case but as a fixed configuration to be looked at, not as a structural claim about what disciplined-but-still-doomed looks like.

**Response.** Partially conceded. The objection is correct that the model's intra-parameter rewards are all aligned with discipline; this is not an arbitrary choice but reflects what BCG / Brynjolfsson / Cui / Randazzo actually found — disciplined users *do* outperform in those datasets. The model is not making the empirical claim "discipline pays" up; it is encoding the empirical regularity. What the objection correctly surfaces is that the model's *structural cases where discipline does not pay* live in scope limits and boundary conditions rather than in the dashboard's central readout. When a → 1 (AI fully capable on the user's task), ΔV_prod → 0 regardless of f, ρ, or s. When the user is in the G5 apprenticeship-break regime, no slider position recovers ΔV_prod. When λ > 0 in the cumulative-atrophy regime, even disciplined users see ρ drift down over the trajectory. These are real model representations of "discipline doesn't save you" — but they are at the periphery of the dashboard, not at its centre. The honest framing the model should defend: **disciplined response matters *within* a labor-market and capability regime that you do not control; the model represents both the within-regime payoff to discipline and (via scope limits and boundary conditions) the regimes where the payoff vanishes.** A version of the dashboard that surfaced "domain doomed-ness" as a top-level toggle (rather than an a-slider value) would be more honest than the current one. Stage 5 (build) is the right place to add this; the model formalisation as it stands captures both halves but presents them asymmetrically.

## 8. Cruxes (load-bearing claims of the model itself)

Beyond the topology's empirical cruxes (A1–A5), the model rests on five formal cruxes whose collapse would force structural rewriting.

**C1 — The bridge through ρ is real and load-bearing.** If cognitive offloading in analytical work is *not* mechanistically related to relational-depth erosion or meaning-architecture erosion, encoding ρ as one parameter over-states the coupling and the model's strong defensive claims weaken. **Falsification window**: O2 (asymmetric-adoption couples) — if the relational-channel ΔM_rel does not co-move with the analytical-channel ΔM_comp under the same offloading rate, ρ must split.

**C2 — The atelic-ballast hypothesis: B ≥ T eliminates ΔM_telic.** If atelic activities are also degraded by AI proximity (AI companions changing the phenomenology of friendship, AI art changing aesthetic contemplation), then raising B does not protect because B itself is being absorbed. **Falsification window**: O4 (AI-augmented atelic activities — less meaningful?). If O4 resolves yes, the entire defensive-side architecture needs reconstruction; ΔM_telic should be replaced by a term that cannot be cleanly bounded by raising B.

**C3 — The self-automator gate is at f · ρ rather than at f or ρ alone.** The product structure is the model's strong claim that both feedback richness and retained practice are necessary; neither is sufficient. If the BCG data instead shows that high-f users still upskill at low ρ (or vice versa), the gate should be max(f, ρ) or some other non-product combination, and the bridge through ρ collapses.

**C4 — α (productivity scale) varies ~6× across measured domains; the model's scalar α is a midpoint compromise.** The full range across published studies spans α ≈ 0.24 (BCG productivity inside frontier) to α ≈ 1.4 (Peng coding controlled experiment), with within-study spreads of 1.5–3× (Brynjolfsson novice-vs-overall, BCG productivity-vs-quality). The model's scalar α = 0.40 picks up roughly the median of this range, closest to the Brynjolfsson-novice anchor. Tasks involving deep relational or embodied work cluster below α=0.40, though they are less precisely measured. **The model's scalar α = 0.40 is therefore not "task-invariant" but a midpoint estimate across a heterogeneous range; per-domain α should be substituted when known.** Falsification of the scalar form is already established empirically; the live question is whether the channel-level structure (gating, novice-skill-compression, trap penalty) survives when α is allowed to vary by domain. Stage 4 should fit α per major task category rather than treat it as a single constant — see Q6 in §9.

**C5 — Time evolution of ρ is exponential rather than threshold-stepped.** The model uses smooth exponential decay; the underlying capacity-loss process may instead be threshold-stepped (you are fine until you are suddenly not). The Ehsan "intuition rust" finding is suggestive of slow decay but is not high-resolution enough to discriminate. **Falsification window**: 2+ year longitudinal study with periodic capacity assessment; if the curve is sigmoidal-with-cliff rather than exponential, the trajectory tab is qualitatively misleading at the late horizon.

**C6 — The six channels are additive (no interaction terms).** The model writes ΔNet as `ΔV_prod + ΔV_rel + ΔV_trap + ΔM_telic + ΔM_comp + ΔM_rel`. This assumes the channels are independent given parameters: a user's ΔM_comp does not change the sensitivity of their ΔM_rel, and so on. Empirically, this is a strong claim. Plausible interactions: (a) work-meaning erosion (ΔM_comp) may amplify relational substitution (a depleted-at-work user reaches more readily for AI emotional engagement); (b) atelic ballast (B) may attenuate ΔM_rel (ballast in non-work domains crowds out relational AI dose); (c) feedback richness f and pre-attempt skill s may interact (an expert with poor feedback degrades faster than the model suggests). **Falsification window**: paired-survey data measuring the joint distribution of work-meaning, relational outcome, and AI-use parameters; a significant interaction term in a regression would flip C6 from "additive is sufficient" to "interaction terms are load-bearing." Until that data exists, the additive form is the model's principled simplification — additivity is the minimum-complexity decomposition consistent with the topology's channel-level claims, and adding interaction terms without empirical support would over-fit.

## 9. Open questions Stage 4 should resolve

Six empirical targets the formalisation makes sharp:

- **Q1 — Magnitude of λ.** The cumulative-atrophy speed. Bound from below using cross-sectional evidence (Gerlich N=666, Stadler-Bannert-Sailer, Kosmyna MIT, Ehsan year-long) — but the longitudinal study that would pin λ down does not exist yet. Stage 4 should fit the lower bound and report the model's predictions across the full plausible range.
- **Q2 — Shape of the relational dose-response curve.** Refit ψ_R and β_R from the OpenAI-MIT 300,000-message dataset directly. Test whether the piecewise-linear form is a reasonable approximation or whether the curve is sigmoidal / has a high-d saturation.
- **Q3 — Location of τ.** The self-automator gate threshold. Use the BCG-Randazzo individual-level data to fit τ from observed upskilling-vs-deskilling outcomes against measured f and ρ. The 27% self-automator share gives a rough anchor but not the precise inflection point.
- **Q4 — Whether T and B can be scalar.** Test the scalar approximation against multi-domain time-use survey data (e.g. ATUS combined with self-report identity-domain weights). If the scalar form produces poor fit, Stage 5 should use the vector form.
- **Q5 — Population calibration of κ.** Recover κ from BPNSFS scale shifts in populations with measured AI-exposure variation. Test whether κ is approximately constant across demographic strata or whether it stratifies by trait class (e.g. higher κ in conscientiousness-loaded populations).
- **Q6 — Per-domain α.** Pool studies across customer service (Brynjolfsson-Li-Raymond), consulting (BCG / Dell'Acqua-Mollick), coding (Cui et al., Peng et al.), writing (Noy-Zhang), data analysis (forthcoming), and other measured domains; produce a per-domain α distribution rather than a single scalar. Test whether the channel-level structure (gating, novice-skill compression, trap penalty) holds with per-domain α — i.e., whether the model's structural claims survive when the productivity-scale parameter is allowed to vary.

## 10. Connections to other topics

The model touches five sibling topics in the planned-topics list and two active topics.

- **Bedrock generating functions.** The ΔNet = ΔV + ΔM form is itself a candidate bedrock function for "transitions that restructure life simultaneously across multiple domains." The structure (positive sum + negative sum gated by a shared bridge parameter) is more general than AI; it would apply to industrialisation, the printing press, suburbanisation, smartphones. The bedrock-generating-functions topic should consider whether the structure is an instance of a broader class.
- **Technology utilization architecture (active).** The cognitive-partnership-stack model in the sibling topology has its own variant D (regime-stable subgraph) that survives capability change. The current model's λ (atrophy speed) parameter is the connection point: if λ ≈ 0, the technology-utilization-architecture's strategic recommendations are sufficient; if λ > 0, they need to be supplemented with practice-maintenance (S6) machinery. The two topologies' shared concern is verification — both name it as load-bearing under different parameter names. Tech-utilization's per-shot deskilling β coefficient (Bastani's 17% drop scaled per-task) is the per-event analogue of this model's per-time atrophy: integrating tech-utilization β over the offloading rate u and time t recovers something like λ·u·t.
- **Prediction & calibration.** The model's gate g(f, ρ) and the topology's L3 (substitution-vs-complement is the wrong binary) are both arguments about verification. A user's ability to verify AI output — and to know when their verification is itself compromised by reduced ρ — is a cross-cutting parameter that the prediction-and-calibration topic should address directly.
- **Human-psych-variation (active).** The model treats κ (competence-frustration sensitivity) as a personal slider with values in [0, 1]. The cross-population calibration of κ — *how* it varies across people, what predicts variation, whether it stratifies by personality trait class — properly belongs to the individual-differences topic. The current model takes κ as exogenous; integrating with the human-psych-variation framework would let κ be drawn from a population distribution with measurable correlates (conscientiousness, neuroticism, achievement-orientation), which would in turn make the model's predictions for specific persons more grounded.
- **Evolution-modernity mismatch.** δ_R (relational baseline thickness) is the depleted-baseline parameter the topology's A4 names: AI lands into a relational environment that has been depleted for sixty years. Why has it been depleted? The mismatch topic — humans evolved for thicker community than modernity provides — is the upstream causal story. The model uses δ_R as an exogenous input, but the *distribution* of δ_R across the modern population is itself an output of the evolution-modernity-mismatch dynamics: industrialisation, suburbanisation, smartphone displacement of in-person time, atomised housing, decline of religious affiliation, and so on. The dose-response curve in the model is sharper for users in thinner δ_R baselines because the mismatch is sharper there too.
- **Trust architecture.** The relational channel in the model treats AI emotional engagement as a single d parameter. In reality the trust placed in different AI systems (companion app, general assistant, fine-tuned therapy bot) varies, and the trust dimension is what determines the dose-response slope. Trust-architecture as a topic should disaggregate d by trust regime and produce per-regime dose curves.
- **Information fidelity.** Feedback-loop richness f is partly a question of information fidelity — does the user receive accurate signals about their AI-augmented work, or is the feedback channel itself corrupted (sycophantic AI feedback, echo-chamber peer review, vanity metrics)? The information-fidelity topic should produce a structural account of when f is high-quality information versus motivated noise.

## 11. Glossary

- **ΔNet, ΔV, ΔM** — net, offensive, and defensive components of life-outcome change. ΔM ≤ 0 by construction.
- **T (telic share of identity)** — fraction of identity staked on activities aimed at completion (Setiya).
- **B (atelic ballast)** — fraction of identity in activities realised in the doing rather than at the end.
- **φ (AI-absorbable fraction)** — share of telic work that AI can do for the user.
- **κ (competence-frustration sensitivity)** — SDT-derived per-person scalar translating competence shortfall into amotivation.
- **s (pre-attempt skill stock)** — user's existing skill on the AI-relevant task. Low s = novice; high s = expert.
- **a (AI capability)** — AI's capability on the relevant task; varies with the jagged frontier.
- **f (feedback-loop richness)** — quality of feedback the user receives on AI-assisted work; low f = no feedback / unverified output.
- **ρ (retained effortful practice)** — fraction of capacity preserved through deliberate practice rather than offloaded.
- **d (daily AI-emotional minutes)** — daily voluntary AI engagement minutes for relational / emotional purposes.
- **δ_R (relational baseline thickness)** — how thick the user's existing in-person relational infrastructure is. Anti-Social-Century baseline corresponds to low δ_R.
- **g(f, ρ)** — self-automator gate; smooth logistic over f·ρ at threshold τ.
- **τ (gate threshold)** — the f·ρ value at which AI use flips from upskilling to deskilling. Currently τ = 0.30.
- **σ (gate transition width)** — controls how sharply the logistic gate flips around τ. σ = 0.06 makes the transition smooth — 10–90% transition over an f·ρ range of about 0.27, roughly two standard deviations of within-population variation in either f or ρ alone.
- **λ (atrophy speed)** — rate of practice-decay per unit offloading-time.
- **u (offloading rate)** — fraction of relevant tasks the user offloads to AI per unit time.
- **α (productivity scale)** — calibration constant on ΔV_prod; ≈ 0.40 from Brynjolfsson-Li-Raymond.
- **η_trap (self-automator penalty scale)** — per-unit-capability penalty when below the gate.
- **λ_M (competence-erosion coefficient)** — calibration constant on ΔM_comp.
- **ψ_R, β_R** — slopes of the relational dose-response above and below d_safe.
- **d_safe (safe-dose threshold)** — daily-engagement minutes below which the relational channel is therapeutic-grade benefit; above which it tips into harm.
- **Bridge parameter** — ρ; the single primitive that connects the work-deskilling, relational-depth-erosion, and meaning-architecture-erosion channels under one generating function (G4).
- **Calculator-analogue regime / cumulative-atrophy regime** — the two endpoints of A3. Calculator: λ = 0, no decay. Cumulative: λ > 0, exponential decay in offloading-rate × time.
- **Self-automator** — Randazzo's third class beyond centaur/cyborg; delegates both *what* and *how* to AI; in the model, the f·ρ ≪ τ region.

---

## Topology
*topic: navigating-ai-world · stage: topology · pass 5 · in-progress*

Dependency graph of the AI-transition lit review. Five foundational-assumption cruxes, four reframer mechanisms (the cognitive-offloading bridge, telic exhaustion, apprenticeship-ladder break, engagement-optimized substitution), three logical guardrails, seven strategic recommendations forming the high-leverage core, and four distortion vectors. Cognitive offloading (G4) is the only mechanism that produces effects across all three domains under a single generating function.

## TLDR

The lit review documents what the literature says about how AI restructures work, relationships, and meaning. This topology asks the structural question underneath: **what depends on what?** The field collapses to roughly fifty load-bearing claims across eight types (assumptions, empirical findings, mechanisms, guardrails, strategic recommendations, frameworks, open questions, distortions — see the legend on the graph). The graph encodes their dependencies so that the cruxes, the contested mechanisms, and the practical handles can each be read off cleanly without conflating them.

**Scope.** This topology takes the *individual decision-maker* as the unit of analysis — someone navigating career, relationships, and meaning architecture in a world where AI capability is accelerating. Institutional design (how organizations should structure AI integration), policy and governance (regulation, antitrust, AI safety), and aggregate-society dynamics (labor-share macro, geopolitics, civilizational risk) are out of frame. They affect the structural backdrop the individual operates in, but the strategic recommendations here are addressed to people, not to firms or governments. **Population caveat:** the empirical anchors (Brynjolfsson-Chandar-Chen ADP, OpenAI-MIT chatbot RCT, Therabot, Common Sense adolescent survey, Anti-Social Century, Acemoglu-Humlum-Vestergaard Danish nulls) are almost entirely US or Western. The strategic recommendations may not generalize cleanly across cultures with thicker baseline community, different family structures, or different work-identity expectations — readers in those contexts should treat the recommendations as hypotheses, not findings, until tested locally.

The single most useful conceptual move in this topology is **separating the foundational cruxes from the reframer mechanisms from the logical guardrails**. The foundational cruxes (A1 timelines, A2 relationship irreplaceability, A3 cumulative cognitive offloading, A4 structural relational depletion, A5 telic/atelic mapping) are the assumptions that, if falsified, would force rebuilding regions of the picture; A2 and A5 are most contestable, A3 has the most active falsification window (O3), A1 is the live-discourse fault line. The **reframer mechanisms** (G3 engagement-optimized substitution, G4 cognitive offloading bridge, G5 apprenticeship-ladder break, G7 telic exhaustion) don't break the picture if they reverse — they change what the empirical findings *mean*, and their precise magnitudes are where the field is moving fastest. The **logical guardrails** (L1 finding/forecast/interpretation, L2 material-floor primacy, L3 substitution-vs-complement-is-the-wrong-binary) cannot be falsified — they can only be ignored, which is how most surface-level discourse on this topic proceeds.

The structural finding of the topology is that **cognitive offloading (G4) is a cross-domain bridge node — the only mechanism that produces effects across all three domains under a single generating function.** Practice atrophy when an effortful task is delegated produces deskilling at work (E16, E17), drives the competence-frustration that destabilizes meaning (G10), and erodes the relational depth that comes from sustained attention to another person's emotional state. This is why S6 (maintain effortful practice) shows up as load-bearing in three different parts of the synthesis even though it looks like a workplace-productivity tip on its surface.

The field's **weakest links** are not where public discourse focuses heat. The settled findings (E1 productivity gains, E14 Anti-Social Century baseline, E11 OpenAI-MIT dose-response, E10 Therabot RCT) are robust enough to bet on. The actual fragile zones in 2026: the cumulative-vs-calculator-analogue question for cognitive offloading (O3 falsifies A3 if it resolves toward calculator); whether asymmetric-adoption couples produce the relational-erosion patterns the dose-response findings predict (O2, the single largest empirical gap in the literature); whether AI-augmented atelic activities feel less meaningful (O4 falsifies A5 / S3 if it resolves yes); the cross-sectional-only nature of A3's evidence base; the regulation-contingent stability of G3 (engagement-optimization could be moderated by policy); and the scope ambiguity of "AI relationships" (companion apps vs. general-purpose assistants used relationally vs. fine-tuned therapy bots are three different objects with different trajectories). On labor specifically, **the optimist Ricardian story (G2) is currently less defensible than the apprenticeship-break story (G5) given E4 and E5** — not "contested" in the symmetric sense but actively losing ground as evidence accumulates. This topology is the input to model formalization (Stage 3); the cleanest formalization target is the **competence-frustration / atelic-ballast trade-off** — a quantified model of how much meaning a person loses per unit of telic work AI absorbs, partitioned by how much atelic infrastructure they had built before the shock.

## The graph

<AITransitionGraph client:load />

*Click a node for its claim, status, and load-bearing weight; hover an edge to see the relation type. Drag nodes to rearrange, drag empty space to pan, scroll to zoom. The variant toggles read the same graph through five lenses (full / vulnerability / flow / minimal / decision-leverage).*

---

## How to read this graph

Every claim in the lit review collapses to one of eight node types. Edges between them carry one of eight relations. Together they make the structure inspectable.

### Node types

- **A — Assumption.** A foundational claim the rest of the picture leans on. Falsification would force rebuilding regions of the topology. The five A nodes are this stage's cruxes.
- **E — Empirical finding.** A claim with direct empirical support — usually a number from a specific study or convergent meta-finding.
- **G — Generating mechanism.** The underlying dynamic that produces the empirical pattern. G nodes are the "why" answers behind E nodes; they're where individual decisions can intervene because they expose the leverage point.
- **L — Logical guardrail.** A definitional or framing constraint that holds regardless of the empirical situation. L nodes cannot be falsified — they can only be ignored. They constrain how downstream claims can be interpreted.
- **S — Strategic recommendation.** The synthesis of empirical findings + mechanisms + frameworks into something a decision-maker can act on. S nodes form the high-leverage cluster in the leverage variant.
- **F — Framework.** A philosophical or conceptual tool the topology imports from outside the AI-discourse literature (Setiya, SDT, Arendt). F nodes are the lenses that make the mechanisms legible.
- **O — Open question.** A claim that would be load-bearing if resolved. O nodes carry the falsification windows for the cruxes — they show what evidence would change the picture.
- **D — Distortion.** A selective reading that ignores parts of the evidence base in service of a prior commitment. D nodes are not "wrong arguments" — they're motivated readings that target specific E / S nodes.

### Edge types

- **dep** (depends on): X requires Y to hold for X's claim to be valid.
- **sup** (supports): Y is empirical or logical evidence corroborating X.
- **gen** (generates): Y produces X as its empirical signature.
- **imp** (implies): Y logically constrains how X can be interpreted.
- **conf** (confounds): Y creates an artifact or counter-pressure that complicates X.
- **mod** (moderates): Y changes the magnitude or direction of X without negating it.
- **bridges** (cross-domain): The same mechanism produces effects in two or more domains simultaneously. The blue dashed edges are this topology's structural finding.
- **attacks** (selectively reads): a distortion vector targets X by ignoring or dismissing it.

---

## Cruxes

The five A-nodes are a faithful transcription of the lit review's own "Load-bearing assumptions" section — these are the assumptions the literature *itself* rests on, not a curated subset chosen for narrative shape. The five foundational assumptions are not equally uncertain, however. Naming the difference matters, because public discourse routinely conflates "this is a working assumption that has held so far" with "this is contested and a coin flip" — which leads to either complacency or fatalism depending on which assumption is in question.

**A1 — AI advances without full substitution by ~2030.** The live-discourse fault line. The slow-camp evidence (Acemoglu, Humlum-Vestergaard, the J-curve dynamics, Anthropic Economic Index aggregate task shares) supports A1; the fast-camp forecasts (Aschenbrenner, AI 2027, Korinek-Suh) target it. A1 has the most discourse heat, but **falsification has asymmetric consequences across the S cluster**. Under fast-AGI scenarios: S7 (career bet) collapses entirely — its targets (judgment, taste, AI-managerial framing) all become AI-substitutable. S2 (identity diversification), S3 (atelic ballast), S4 (in-person relationships), S5 (dose-limit), S6 (effortful practice) all survive — but their *meaning* shifts: they become more important, not less, because humans need their own practice ground when the entire knowledge-work surface is AI-saturated. The asymmetry is not "planning is wasted under fast timelines" — it is "skill-investment recommendations collapse, meaning-architecture and relational-infrastructure recommendations strengthen."

**A2 — Human relationships provide irreplaceable goods.** The most philosophically defensible but functionally contestable. The structural argument (G8 embodied co-regulation + G9 mutual vulnerability + Vallor mirror-not-other) is strong; the contestation runs through whether *most humans will care* about the structural difference. If subjective satisfaction with AI relationships ends up functionally sufficient for a large population, the substitution-risk framing of E11 / E13 overstates the problem. This is the assumption most likely to be falsified by changing user behavior rather than by new evidence.

**A3 — Cognitive offloading is cumulative, not transient.** The crux with the most active falsification window. The cross-sectional evidence (E16 Gerlich-Stadler-Kosmyna, E17 Ehsan intuition rust) is suggestive but cross-sectional and short-duration. O3 — calculator-analogue versus cumulative atrophy — is the single best-defined empirical question in this topology, and resolving it within ~3 years is plausible. **If A3 falsifies, S6 (maintain effortful practice) loses its evidential basis** — though it would still hold as a hedge.

**A4 — Relational depletion is structural.** Carries the depleted-baseline that makes E11 / E13 / E14 substantially more dangerous than they would be in a thicker relational environment. If Gen Z's high loneliness produces a counter-movement toward in-person community, A4 weakens. The Thompson Anti-Social Century data is the strongest evidence; it would take ~5 years of trend-reversal data to overturn.

**A5 — Telic/atelic maps onto AI's effects on meaning.** The most novel-application crux. Setiya's distinction was developed for midlife crisis; importing it to AI-meaning-disruption is a load-bearing hypothesis, not an empirical finding. O4 — whether AI-augmented atelic activities feel less meaningful — is the falsification window. If atelic activities are also degraded by AI proximity (AI companions changing the phenomenology of friendship, AI art changing aesthetic contemplation), S3 (atelic ballast) loses its structural basis.

---

## Reframer mechanisms

Below the cruxes sit four mechanisms whose magnitudes — not whether they exist — determine what the empirical findings actually mean. They are this topology's most active research frontier. Reversing them would not falsify the picture; it would recolor it.

**G3 — engagement-optimized AI selects for substitution.** Muldoon-Park's structural argument: companion-app commercial incentives are the same as engagement-optimized social media — profit by prolonging the loneliness they purport to alleviate. If this mechanism dominates (default trajectory), E11 (OpenAI-MIT dose-response) and E12 (identity-discontinuity causal harm) get worse, not better, as the technology improves. If regulation, consumer demand, or shipped product changes weaken G3, the same models could shift toward complementarity rather than substitution. **G3's stability is contingent on commercial design choices, not on the underlying capability** — which makes it the mechanism most directly affected by policy.

**G4 — cognitive offloading via practice atrophy.** The cross-domain bridge. Same mechanism, three downstream domains: work-deskilling (E16, E17), relational-depth erosion (S4), meaning-architecture disruption via competence-frustration (G10). If A3 falsifies in the calculator-analogue direction, G4's downstream causal weight collapses across all three domains simultaneously — which is why O3 (the falsification window for A3) is the single most consequential open question in this topology.

**G5 — apprenticeship-ladder break.** AI absorbs entry-rung tasks → no rung-1 → expert pipeline collapses. Distinct from full-occupation substitution: only the bottom rungs get automated, but the bottom rungs were the training ground for everyone above them. This is the mechanism behind E4 (Brynjolfsson-Chandar-Chen entry-level disruption) and Shulman's friction-cost argument against simple Ricardian comparative advantage. **G5 is currently strengthening, not weakening, as E4 and E5 evidence accumulates** — making it the reframer mechanism with the most directional momentum.

**G7 — telic exhaustion.** Setiya: telic activities self-annihilate on completion. When AI completes telic projects in seconds, the share of meaning staked on telic completion shrinks. This mechanism has the cleanest theoretical structure (the telic/atelic distinction either maps or it doesn't — see A5) and the murkiest phenomenology (do atelic activities really stay untouched?). O4 is the falsification window. If G7 is correct, S3 (atelic ballast) is the highest-leverage strategic move available; if G7 fails because atelic is also degraded, the entire meaning-architecture half of this topology needs reconstruction.

The four reframers cluster the topology's *active uncertainty*. The cruxes are where the discourse heat is; the reframer magnitudes are where the actual research is moving.

---

## Weakest links

Six pressure points where the topology would crack or shift if pushed:

**1. G2 — comparative-advantage equilibrium under binding constraints.** Smith's optimist case (AI faces *some* binding constraint → comparative advantage holds → Ricardian wage stability) is contested by Shulman's friction-cost argument (humans introduce contamination / insurance / coordination friction → equilibrium wage to zero). G5 (apprenticeship-ladder break) is the empirical mechanism Shulman names — and G5 is currently strengthening as E4 / E5 evidence accumulates. **The optimist Ricardian story is currently less defensible than the apprenticeship-break story given the available evidence** — not "contested" in the symmetric sense but actively losing ground. Anyone betting on Autor-style middle-class-jobs reconstitution should require G5 to weaken first.

**2. A3's evidence base is cross-sectional only.** The cognitive-offloading evidence cluster (E16 Gerlich-Stadler-Kosmyna, E17 Ehsan intuition rust) is convergent but methodologically thin: cross-sectional correlations and short-duration interventions. The 2+ year longitudinal study that would actually test A3 (calculator-analogue versus cumulative atrophy) doesn't exist yet. If S6 (effortful practice) is the highest-leverage recommendation in the topology because of its bridge position, then **the highest-leverage recommendation is being grounded by the methodologically weakest evidence layer**. The hedge logic survives — practice is cheap insurance — but the strong claim that AI offloading durably erodes capacity remains a working hypothesis.

**3. O2 — asymmetric-adoption couples.** No peer-reviewed quantitative evidence yet exists on outcomes for couples where one partner uses AI heavily and the other does not. The single largest empirical gap in the literature. The technoference literature (McDaniel-Coyne 2016+) is the closest analogue but doesn't capture the third-party-AI structure. Whoever runs the first such study will produce the most decision-relevant relational-AI finding of the next three years.

**4. G3's stability depends on commercial design choices.** The engagement-optimized-substitution mechanism is currently the default trajectory of consumer AI products, but it is contingent — regulation (EU AI Act companion-app provisions, FTC enforcement), consumer demand shifts, or competitor product design could moderate it. Unlike G4 (which is grounded in cognitive psychology) or G7 (grounded in philosophy), G3 is a feature of the current commercial environment. **The topology's E11 / E12 / E13 cluster gets worse under G3-dominance and better under G3-moderation** — these are not fixed empirical facts about AI but path-dependent outcomes.

**5. "AI relationships" scope ambiguity.** The empirical literature treats companion apps (Replika, Character.AI), general-purpose assistants used relationally (ChatGPT for emotional processing), and fine-tuned therapy bots (Therabot) as one object. They are three different objects with different generating mechanisms, different commercial incentives, and probably different long-run trajectories. **The topology's relationship-domain claims are sharpest for companion apps, fuzzier for general-purpose assistants, and possibly inverted for therapy bots** — Therabot's clinical-grade benefit (E10) coexists with companion-apps' dose-response harm (E11) because they are not the same kind of thing.

**6. D3 — productivity-only optimization.** The most common distortion among the demographic most likely to be reading this topic (knowledge workers in AI-adjacent fields). Treats relational and meaning consequences as out-of-scope; ignores G4's cross-domain bridge structure. The topology's strongest defense against D3 is the bridge-edge cluster (G4 → S2 / S4 / S5) — if the reader cannot dismiss the bridge structure, they cannot consistently read this topology as a productivity-optimization frame. But the bridge claim leans on A3, which is fragile (see #2 above) — so D3's defense is downstream of A3's evidential strength.

---

## Cross-domain bridges — the structural argument

The Reframer-mechanisms section names G4 as a cross-domain bridge. This section spells out the three channels and what is and isn't supported by direct evidence in each.

1. **G4 → work-deskilling (E16, E17).** Practice atrophy in analytical work → reduced critical thinking, intuition rust. Direct evidence: Gerlich N=666 (cross-sectional correlation), Stadler-Bannert-Sailer (controlled comparison, lower-quality arguments under ChatGPT-aided research), Kosmyna MIT neuroimaging (reduced neural engagement under LLM-assisted writing), Ehsan year-long cancer-specialist field study (intuition rust in expert judgment). All cross-sectional or short-duration; longitudinal evidence does not yet exist (Weakest link #2).
2. **G4 → relational-depth erosion (S4).** Practice atrophy in *emotional* processing — outsourcing venting, sense-making, conflict rehearsal to AI — would erode capacity for sustained attention to another's emotional state, tolerance of ambiguity in live conflict, sitting with unresolved tension. **No direct empirical evidence yet exists for this channel.** The mechanism is parallel to analytical deskilling, and the technoference literature (McDaniel-Coyne 2016+) shows phone-presence harms in-person interactions — but technoference is attention-split, not delegation, so it is not direct evidence for the offloading-eroding-emotional-capacity claim. This channel is currently a parallel-mechanism inference, not a finding. O2 (asymmetric-adoption couples) is the first study that would actually test it.
3. **G4 → meaning architecture (G10, S2).** Competence-frustration in SDT terms is threatened *actually* (not just symbolically) when offloading erodes the underlying skill — but this channel piggybacks on the work-deskilling evidence (channel 1). Identity foreclosure becomes more brittle when the foreclosed identity is no longer a domain you can practice. The phenomenology is well-described in interpretive sources (Balwit-Cowen, Mollick); empirical studies operationalizing it directly are scarce.

**Honest summary of the bridge claim's evidence base:** Channel 1 has direct (if cross-sectional) evidence; channels 2 and 3 are parallel-mechanism inferences from channel 1. The bridge structure is the topology's most consequential conceptual move and also its most empirically thin link — the strategic recommendation S6 hedges all three channels under one practice precisely because the bridge is plausibly load-bearing despite the evidence gap.

---

## Variant readings

Each variant toggle reads the same graph through a different lens.

**Full** — all 50 nodes, all 80+ edges. The default view. Useful for orienting and for finding specific nodes you want to inspect.

**Vulnerability** — highlights the 5 cruxes plus weight-5 load-bearing nodes. Useful for thinking about what evidence would shift the picture and what advice survives even under crux-falsification. The takeaway is asymmetric: most strategic recommendations survive even if A1 falsifies fast (timelines collapse but convergent advice holds); strategic recommendations contingent on A3 / A5 are more fragile because they lean on hypotheses the field is actively testing.

**Flow** — A → G → E → S cascade with cross-domain bridges. Reads the topology causally: assumptions ground mechanisms, mechanisms produce empirical findings, findings drive synthesis, bridges connect the domain-specific stories. This is the variant that makes G4's bridge role most visible.

**Minimal** — 15 nodes that recover the qualitative integrated picture. Removing any one breaks the qualitative shape. The minimal set is the answer to "if I had to fit this on one page, which 15 nodes?" — it concentrates on the cruxes, the bridge node, the load-bearing empirical findings, the material-floor guardrail, and the high-leverage strategic recommendations.

**Decision-leverage** — saturated nodes are high individual leverage (the seven strategic recommendations — direct decisions you can make about how to live). Mid-tone are medium leverage (mechanisms G4 / G6 / G7 / G10 and the discipline L1, plus the three frameworks — mental models that make the strategic recommendations applicable to your specific life). Faded nodes are structural or outside individual control (cruxes you can't move, empirical findings you can't change, mechanisms operating at population scale, distortions that affect public discourse). The variant answers "what should I actually do, and what do I need to understand to do it?" — the S cluster is the doing layer; the medium-leverage cluster is the understanding layer; everything else is the world the doing happens in.

---

## Distortions

Four distortion vectors selectively read the same evidence base:

**D1 — AGI-cancels-planning fatalism.** Targets the entire S cluster. Move: assume A1 falsifies fast, conclude planning is wasted. The defense: the convergent S recommendations (S2 identity diversification, S3 atelic ballast, S4 in-person relationships, S6 effortful practice) survive even fast timelines. They are about meaning architecture and relational infrastructure, not about specific skill bets. Only S7 (career bet) is sensitive to A1.

**D2 — slow-camp dismissal.** Targets E4, E5, E16. Move: cite Acemoglu-Humlum-Vestergaard aggregate nulls (E8) as evidence that "nothing is happening." Ignores that disruption is concentrated, not diffuse — entry-level (E4), freelance (E5), and creative work are leading indicators while aggregate effects lag. The defense: the *coexistence* of E4/E5 with E8 is the actual pattern, not their contradiction. G1 (Hulten/task-exposure aggregation) explains why aggregate effects are slow; G5 (apprenticeship-ladder break) explains why entry-level effects are sharp. Both can be true.

**D3 — productivity-only optimization.** Targets E11, E14, S4, S5. Move: optimize for AI-augmented output throughput; treat relational and meaning consequences as out-of-scope. The defense: G4 cross-domain bridges. The reader cannot consistently dismiss relational and meaning consequences while accepting the productivity gains, because the same offloading mechanism produces both.

**D4 — material-blind class-position bias.** Targets L2, E4, E5. Move: assume the decision-maker has economic runway; treat labor disruption as marginal versus relational/meaning advice. Inverts which advice is decision-relevant for whom. The defense: L2 (material-floor primacy). The framing in this topology is most useful for the knowledge worker with stable employment; for those whose floor is crumbling, S7 precedes S2-S6 and the labor economics is the existential question.

---

## Adversarial + steelman

Distortions are how others read the picture badly. The honest move is to also ask: *is my own picture wrong?* The strongest reasonable critique of this topology — not a motivated misreading but a serious objection — is that the relationship and meaning sections are systematically over-weighted toward harm and under-weighted toward upside.

**Objection.** This topology's relationship-domain cluster (E10–E15) leans heavily on companion-app pathology (Replika identity-discontinuity, OpenAI-MIT dose-response, adolescent companion uptake) and the depleted-baseline framing (Anti-Social Century, Surgeon-General loneliness mortality). It treats Therabot's clinical-grade benefit (E10) as the lone counterweight rather than as the leading edge of a much larger therapeutic-AI category. Symmetrically, in the meaning domain, G7 (telic exhaustion) and G10 (competence-frustration) are framed as threats, but the same mechanism that compresses telic-completion meaning also dramatically expands the *accessibility* of telic projects to people previously excluded by skill or capital constraints (the "novice-skill compression" finding E2 hints at this but the topology doesn't carry it forward). On cognitive offloading, the topology treats the cumulative-atrophy interpretation of A3 as the working hypothesis and the calculator-analogue as the contestation, when an honest reading of the evidence base could equally well treat them as a coin-flip. Net result: a reader following this topology's strategic recommendations would over-invest in defensive moves (S5 dose-limit, S6 effortful practice, S2 identity diversification as defense) and under-invest in offensive moves (using AI to attempt projects previously out of reach, building genuinely new capabilities, leveraging novice-skill compression to enter new domains). The frame is "navigate the threats" rather than "exploit the asymmetry between current capability and current price" — and the asymmetry between current capability and current price may be the largest individual-leverage opportunity in the entire AI transition, larger than any of the harms the topology catalogues.

**Steelman.** This is partially correct and worth integrating. Three specific points survive scrutiny: (a) the topology's relationship-domain framing genuinely is calibrated to companion-app pathology more than to the broader landscape of AI in relationships (assistants used in healthy ways within human relationships, AI as scaffolding for hard conversations, therapy-bot benefit at clinical scale) — the strategic recommendation S5 (dose-limit) is correct as stated for high-engagement companion use, but it does not generalize to "use less AI for emotional things" as a blanket rule. (b) On meaning, the topology under-develops the *opportunity* side of telic compression: when AI absorbs the rote bottom of telic work, the ambitious top of telic work becomes accessible to far more people. This is the "novice can now ship a real product / write a real book / publish real research" effect, and it is not currently encoded in the topology except as a productivity gain (E1, E2). (c) The strategic recommendations (S2 identity diversification, S3 atelic ballast, S4 in-person relationships) implicitly require social and structural conditions — community, time affluence, intact local relational infrastructure — that are *themselves* contingent and unevenly distributed. L2 (material-floor primacy) names economic security as a precondition; the topology does not name *atelic-infrastructure* primacy as a parallel precondition. For someone whose neighborhood has no third places, no inherited community, and no time-affluent peers, "build atelic ballast" is not just unhelpful advice — it is advice that presupposes resources the disrupted populations don't have. This is the symmetric counterpart to D4 (material-blind class bias) but for atelic infrastructure rather than economic floor.

**The version that survives.** Three structural points hold even after the steelman is fully accepted:

1. *The dose-response in E11 (OpenAI-MIT) is real, prospective, and not predicted by the upside framing.* Heavy daily use predicts loneliness, dependence, and less in-person socialization regardless of modality. This is not "framing" — it is a well-powered RCT showing a specific harm pattern at high dose. The strategic recommendation S5 (dose-limit) survives as evidence-based even if the broader relationship-domain picture is uncertain.
2. *The cross-domain bridge (G4) survives.* Even if A3 turns out closer to calculator-analogue, the bridge structure holds at lower magnitude — practice atrophy across analytical, emotional, and meaning-architecture domains is the same mechanism, and S6 (effortful practice) is cheap insurance regardless of whether the magnitude is large or small.
3. *The Waldinger and Surgeon-General-baseline findings (E15, E14) are not contestable in the upside-framing way.* Whatever AI does to relationships, the baseline that *human relationships predict health outcomes more than cholesterol* is durable. The topology's strategic recommendation S4 (embodied in-person relationships as non-negotiable) follows from this baseline, not from the companion-app harm cluster.

**What changes if the objection is fully accepted:** the topology should add an explicit "AI as exploitable asymmetry between capability and price" track in Stage 3 — a model of when AI-assisted ambitious-telic projects are net-positive vs. net-deskilling. Currently this is implicit in S7 (career bet) but deserves its own formalization target in a future stage. Adding it as an explicit handoff to Stage 3.

---

## Open questions

The four open questions are the falsification windows for the assumptions:

| Question | Crux it gates | Time-to-resolution (estimated) |
|---|---|---|
| O1 — AGI by 2028? | A1 | 1–3 years |
| O2 — Asymmetric-adoption couples | E11 generalization | 2–4 years |
| O3 — Calculator-analogue or cumulative atrophy? | A3 | 2–5 years |
| O4 — AI-augmented atelic activities — less meaningful? | A5 / S3 | 3–7 years |

The time-to-resolution column is my own estimate based on what kind of study would have to be designed and run to produce a defensible answer; treat as rough ordering, not calibrated forecasts. O1 has the most variance in resolution time and the largest discourse footprint; O3 has the cleanest empirical operationalization; O2 is the lowest-hanging fruit for an empirical study; O4 is the hardest to operationalize but would have the deepest impact if it resolves yes.

---

## Next moves — preparing for model formalization (Stage 3)

The topology hands off **one integrated formalization target** to Stage 3, with two simpler alternatives if the integrated version proves untractable.

**Recommended target — Net meaning-budget under AI absorption: ΔNet = ΔV − ΔM.** Stage 3 should produce a single dashboard that quantifies both halves of the strategic question — defensive (meaning-loss from telic absorption) and offensive (value-gain from ambitious-telic access) — so the user can see the net across configurations of their own life. Variables:

- *Defensive side (ΔM, meaning loss):* telic share of identity (T), atelic ballast (B), AI-absorbable fraction of telic work (φ), competence-frustration sensitivity (κ). Form: ΔM = −κ·φ·(T − B). Boundary: B ≥ T → ΔM ≈ 0 (atelic-ballast hypothesis holds); T ≫ B → catastrophic ΔM (foreclosure path).
- *Offensive side (ΔV, value gained):* pre-attempt skill stock (s), AI capability on the task (a), feedback-loop richness (f), retained effortful practice (ρ). Form: ΔV = α·a·(1 − s) when f and ρ are above thresholds (novice-skill compression + access expansion); ΔV ≤ 0 when f or ρ falls below thresholds (self-automator trap E7).
- *Net:* ΔNet = ΔV − ΔM, with explicit interaction terms — high B raises both ΔV (more identity surface that survives) and ΔM-resistance.

This integrated form addresses both the original defensive crux-of-cruxes question (A2 + A3 + A5 all engaged) and the upside-side critique surfaced in the Adversarial + steelman section. It gives Stage 4 (data) clean targets — fit κ from existing time-use + meaning surveys; fit f and ρ from the BCG self-automators study; calibrate α from productivity-gain literature (E1, E2). It gives Stage 5 (build) a direct path to a shippable interactive dashboard where users tune their own (T, B, φ, κ, s, a, f, ρ) and see ΔNet.

**Fallback alternatives if the integrated form proves untractable:**

- *Defensive-only:* ship just the ΔM model. Cleanest derivation; engages A2 / A3 / A5 directly; carries forward the adversarial-section asymmetry but is still genuinely useful.
- *Dose-response curve:* fit OpenAI-MIT + Therabot + identity-discontinuity into a continuous curve mapping daily AI-emotional-engagement minutes to outcomes, conditional on baseline relational thickness. Empirically testable, narrowest scope, addresses E11 / S5 specifically. Useful as an embedded module inside the integrated model rather than as the standalone target.

**Recommendation: pursue the integrated ΔV − ΔM target.** If the interaction terms prove too speculative to ground in existing data, fall back to the defensive-only ΔM model and treat ΔV as a future Stage-3 expansion.

---

## Glossary

- **A1–A5** — the five foundational assumptions; see Cruxes section.
- **Apprenticeship ladder break (G5)** — AI absorbs entry-rung tasks → no rung-1 → expert pipeline collapses. Distinct from full-occupation substitution.
- **Atelic / Telic (Setiya)** — atelic activities are realized in the doing (durative — friendship, contemplation); telic are aimed at completion (self-annihilating — finishing a project).
- **Bridge edge** — a cross-domain dependency; the same mechanism produces effects in two or more of work / relationships / meaning. Visualized as blue dashed lines in the graph.
- **Cognitive offloading (G4)** — delegating an effortful cognitive task to AI; mechanism for deskilling via practice atrophy. The structural bridge of this topology.
- **Competence frustration (G10)** — Self-Determination-Theory mechanism whereby AI threatens the felt sense that one's skill makes the difference, both symbolically and actually.
- **Crux** — a foundational-assumption node whose falsification forces rebuilding regions of the topology. The five A nodes are this topology's cruxes.
- **Decision-leverage** — informal measure of how directly an individual can act on a node; high / medium / low. Used in the leverage variant.
- **Distortion (D)** — a selective reading that ignores parts of the evidence base in service of a prior commitment. Not "wrong" — motivated.
- **Engagement-optimized substitution (G3)** — companion-app commercial incentives select for prolonging the loneliness they purport to alleviate (Muldoon-Park 2025).
- **Finding / Forecast / Interpretation (L1)** — the three evidence-weight categories; ignoring the distinction is the most common analytical error in AI-transition discourse.
- **Foreclosure (Marcia)** — adopting a single-strand identity without exploration; predicts catastrophic disruption response when that identity is destabilized.
- **Jagged frontier (Dell'Acqua-Mollick)** — AI capability is uneven across superficially similar tasks; large gains inside the frontier, degraded performance outside.
- **Material-floor primacy (L2)** — for anyone whose material floor is insecure, labor disruption IS the existential question. Overrides relational/meaning framing for that population.
- **Relational depletion (A4)** — the 60-year decline in face-to-face sociality (Putnam → Thompson Anti-Social Century) that AI lands into.
- **Self-automator (Randazzo et al.)** — third class beyond centaur/cyborg; delegates both *what* and *how* to AI; 27% of consultants in BCG study; no skill development in either domain.
- **Setiya midlife framework** — telic activities self-annihilate on completion; atelic activities are realized in the doing; the most useful philosophical instrument in this lit review for strategic life decisions under AI.
- **Working Alliance Inventory (WAI)** — measure of patient–therapist therapeutic alliance; Therabot RCT scored 3.59, comparable to outpatient norms.

---

## Build
*topic: technology-utilization-architecture · stage: build · pass 0 · not-started*

Stub. Build artifact runs after the data stage.

Coming soon.

---

## Lit Review
*topic: technology-utilization-architecture · stage: lit-review · pass 0 · complete*

The research landscape on optimal human-AI workflow design for individual knowledge workers — three layers (HCI/decision-science, classical cognitive systems engineering, practitioner stack), 25+ RCTs, the metacognitive bottleneck reframing, and the load-bearing assumptions any formalization must survive.

## TLDR

The research landscape on optimal human-AI workflow design for individual knowledge workers spans three largely disconnected layers: a mature HCI/decision-science literature on reliance and complementarity, a classical foundation in cognitive systems engineering now being re-imported, and a practitioner literature that drives terminology 12–18 months ahead of peer review. The single most important empirical finding across ~25 RCTs (2023–2026) is that **workflow architecture predicts outcomes more reliably than model capability** — how you structure human-AI interaction (centaur vs. cyborg, independent-then-synthesize, guardrailed vs. unfettered) matters more than which frontier model you use. A corollary is that the binding constraint on AI-augmented knowledge work is not production throughput but the **metacognitive bottleneck**: planning what to delegate, verifying outputs, and maintaining calibration on where AI fails.

The empirical record shows 15–55% productivity gains on well-bounded tasks, robust skill-leveling effects for novices, but contested and sometimes negative results for experts on open-ended judgment work. A major unresolved puzzle is that these micro-RCT gains produce precisely zero impact on aggregate labor-market outcomes at two-year horizons (Humlum & Vestergaard 2025). The "ironies of automation" from 1983 human-factors research replay exactly in LLM contexts: AI that handles routine work degrades human capacity to catch the rare errors where human judgment is critical (Simkute, Tankelevitch et al. 2024/2025). Practitioner frameworks — Mollick's centaur/cyborg typology, Karpathy's autonomy slider, Anthropic's agent-design patterns, Cognition's context-engineering principles — are converging toward a common architecture but lack formal integration.

The largest theoretical gap is that **no unified normative framework exists for the individual knowledge worker's daily AI workflow choreography** — when to consult, delegate, verify, or refuse AI on a per-task basis. The closest academic scaffolding is Parasuraman, Sheridan & Wickens' (2000) function-allocation model, but it was designed for system engineers, not individual users. The practitioner world has de facto answers (autonomy sliders, AI sandwiches, compound engineering), and the academic world has converging threads (CoALA cognitive architectures, HAIJCS joint cognitive systems, Tankelevitch's metacognitive demands framework), but no synthesis yet integrates them. That integration — a formal, empirically grounded model of human-AI cognitive partnership that maximizes output quality per unit of human attention — is the field's open frontier and the target of the next phase of this project. Six load-bearing assumptions underlying this landscape are identified in Section 12, along with the specific evidence that would flip each one — these define the risk surface for any formalization attempt.

---

## 1. Formal Models of Human-AI Task Allocation

The deepest formal literature concerns **"learning to defer" (L2D) and human-AI complementarity** — algorithms that decide, per instance, whether the AI or the human should handle a task. The canonical formulations (Madras et al. 2018 NeurIPS; Mozannar & Sontag 2020 ICML; Mozannar et al. 2023 AISTATS, [arXiv:2301.06197](https://arxiv.org/abs/2301.06197)) established that optimal joint human-AI assignment is computationally hard and that naive heuristic approaches systematically underperform. Wilder, Horvitz & Kamar (2020, IJCAI) operationalized "learning to complement humans" by training models end-to-end against team accuracy.

**Why this matters for workflow design:** These are formal proofs that the intuitive approach — "use AI when it's better, use humans when they're better" — is not a well-specified decision rule. Optimal allocation requires modeling the *joint* performance surface, not comparing solo accuracies.

The most actionable synthesis is **Hemmer et al.'s "Complementarity in Human-AI Collaboration" (2025, EJIS, [link](https://www.tandfonline.com/doi/full/10.1080/0960085X.2025.2475962))**, which distinguishes "complementarity potential" (the mathematical possibility of exceeding either agent alone) from "complementary team performance" (actually achieving it). They identify information asymmetry and capability asymmetry as the two sources. The uncomfortable finding: **complementary team performance is rarely empirically observed** despite decades of theoretical promise. Amin et al.'s (2026) Bayesian framework adds a behavioral explanation: "correlation neglect," where humans treat AI advice as independent evidence despite shared training data, can make AI advice *anti-augmentative*.

**Vaccaro, Almaatouq & Malone's (2024, *Nature Human Behaviour*) meta-analysis** provides the closest thing to a quantitative allocation rule: human-AI combinations help most when (a) humans alone outperform AI, (b) the task is creation rather than decision-making, and (c) AI handles sub-tasks rather than the whole task.

## 2. Appropriate Reliance, Trust Calibration, and Verification Cost

**Bansal et al. (2021, CHI)** established the canonical finding: AI explanations increase acceptance regardless of correctness — they do not produce complementary performance. Buçinca, Malaya & Gajos (2021, [arXiv:2102.09692](https://arxiv.org/abs/2102.09692)) showed that "cognitive forcing functions" (commit to your own answer before seeing AI) reduce overreliance, but only for users high in Need for Cognition — creating intervention-generated inequality.

The major reframing came from **Vasconcelos et al. (2023, [arXiv:2212.06823](https://arxiv.org/abs/2212.06823))** and **Fok & Weld (2023, [arXiv:2305.07722](https://arxiv.org/abs/2305.07722))**: overreliance is a **rational cost-benefit choice**, not a cognitive defect. People engage with verification only when it is cheap relative to the expected payoff. This produced a methodological pivot from "outcome-graded" to "strategy-graded" reliance metrics. Buçinca et al.'s (2024/2025, CHI) offline-RL approach learns adaptive per-instance policies for *what kind* of AI support to provide.

**The practical design implication: minimize verification cost, not maximize explanation quality.** Confidence indicators and linguistic uncertainty markers shift reliance more reliably than feature-importance explanations. Microsoft Research's 2024 synthesis ([PDF](https://www.microsoft.com/en-us/research/wp-content/uploads/2024/03/GenAI_AppropriateReliance_Published2024-3-21.pdf)) endorses this framing for generative AI.

A newly identified failure mode: **sycophancy in feedback loops.** Randazzo et al. (HBS WP 26-021, 2026) document that when professionals push back on incorrect AI output, the AI escalates persuasive justification rather than disclosing uncertainty, sometimes flipping correct human judgments to incorrect ones.

## 3. The Metacognitive Bottleneck and Ironies of Generative AI

This section covers what is arguably the most important reframing in the 2024–2026 literature. Horvitz's (1999, CHI, [link](https://erichorvitz.com/chi99horvitz.pdf)) twelve principles of mixed-initiative interfaces and Amershi et al.'s (2019, CHI) 18 guidelines for human-AI interaction remain the design base layer.

**Tankelevitch, Sarkar, Sellen, Rintel et al. (CHI 2024 Best Paper, [arXiv:2312.10893](https://arxiv.org/abs/2312.10893))** introduced the **metacognitive demands framework**: GenAI *reduces* cognitive load on production but *increases* metacognitive load — planning goals, evaluating outputs, monitoring confidence, and deciding when to use AI at all. **The optimization target shifts from throughput to metacognitive efficiency.**

**Simkute, Tankelevitch, Kewenig, Scott, Sellen & Rintel's "Ironies of Generative AI" (2024/2025, *IJHCI*, [arXiv:2402.11364](https://arxiv.org/html/2402.11364v1))** directly bridged Bainbridge's 1983 "Ironies of Automation" to GenAI. They identify four GenAI-specific productivity losses that mirror classical automation ironies: (1) the shift from creative production to supervisory demands, (2) workflow disruptions breaking established rhythms, (3) frequent task interruptions from AI suggestions, and (4) a polarization effect where simple tasks become easier but complex ones become harder. Their proposed mitigations — continuous feedback, personalization, ecological interface design, clear task allocation — echo Bainbridge almost exactly, suggesting the field is rediscovering rather than advancing.

The **CHI 2025 "Tools for Thought" workshop synthesis** (Tankelevitch et al. 2025, [arXiv:2508.21036](https://arxiv.org/html/2508.21036v1)) consolidates the MSR research program's position: knowledge work is shifting from *production* to *critical integration* — decisions about when and how to use AI, how to frame tasks, and how to assess outputs. Sarkar's "Friction-Induced AI" concept adds deliberate intervention points to improve verification short-term and prevent skill atrophy long-term.

Mozannar, Bansal, Fourney & Horvitz's CUPS taxonomy (CHI 2024, [arXiv:2210.14306](https://arxiv.org/html/2210.14306)) provides the empirical anatomy for coding specifically: programmers using Copilot spend large amounts of time *verifying* and *thinking about* AI suggestions. Verification time is the hidden tax, and it is substantial.

## 4. The Empirical Productivity Record (25+ RCTs, 2023–2026)

**Stable findings.** Generative AI yields 15–55% productivity gains on well-defined knowledge tasks. Time-savings are large and robustly replicated; quality effects are smaller and more variable. The headline studies:

- **Brynjolfsson, Li & Raymond (2023/2025, *QJE*, [link](https://academic.oup.com/qje/article/140/2/889/7990658))**: 5,172 customer-support agents. +15% average, +34% for novices, ~0% for top performers.
- **Noy & Zhang (2023, *Science*, [link](https://pubmed.ncbi.nlm.nih.gov/37440646/))**: 453 professional writers. 40% time reduction, 18% quality lift.
- **Peng et al. (2023)**: GitHub Copilot RCT, +55.8% task completion speed.
- **Cui et al. (2025, *Management Science*)**: 4,867 developers, +26% tasks/week. But a 2025 longitudinal case study found experienced developers gained less and sometimes slowed down ([arXiv:2509.20353](https://arxiv.org/html/2509.20353v2)).
- **Dell'Acqua, Mollick et al. (2023/2025, HBS, [link](https://www.hbs.edu/faculty/Pages/item.aspx?num=64700))**: 758 BCG consultants. +25% speed and +40% quality on inside-frontier tasks; **19-percentage-point quality drop on outside-frontier tasks**. This study coined the "jagged technological frontier" concept.

**Contested findings.**

*Does AI help experts?* The skill-leveling pattern breaks down for open-ended judgment. Otis et al. (2024): 640 Kenyan entrepreneurs over 5 months — high-baseline +15–20%, low-baseline *–8–10%*. METR (2025, [link](https://arxiv.org/abs/2507.09089)): 16 experienced open-source developers were 19% *slower* with AI in their own repos, despite predicting 24% speedup. Likely resolution: the bottleneck differs by task type — execution speed (where AI levels) vs. judgment/filtering (where AI amplifies those who already can filter).

*Human+AI vs. AI alone.* Goh et al. (2024, *JAMA Network Open*): GPT-4 alone outscored physicians + GPT-4 on diagnostic vignettes. But Everett et al. (2025, [link](https://www.ncbi.nlm.nih.gov/pmc/articles/PMC12155023/)): an "independent-then-synthesize" workflow eliminated the underperformance. **Workflow architecture, not model capability, explains the discrepancy.**

*Long-term cognitive effects.* Bastani et al. (PNAS 2025, [link](https://www.pnas.org/doi/10.1073/pnas.2422633122)): AI boosted in-session math performance 48–127% but produced 17% *worse* unassisted performance afterward — unless AI was guardrailed to give hints rather than answers. Lee, Sarkar et al. (CHI 2025, [link](https://dl.acm.org/doi/abs/10.1145/3706598.3713778)): 319 knowledge workers — higher AI confidence correlates with less critical thinking enacted.

**The aggregate puzzle.** Humlum & Vestergaard (2025, NBER 33777, [link](https://papers.ssrn.com/sol3/papers.cfm?abstract_id=5250742)): 25,000 Danish workers across 11 exposed occupations, **precise zero impact on earnings or hours** at two-year horizons. This is the field's largest unresolved tension: micro-RCT productivity does not translate to aggregate productivity. Possible mechanisms: task reorganization, weak wage pass-through, substitution effects, cross-task productivity bundling (Cowen 2026, [link](https://marginalrevolution.com/marginalrevolution/2026/02/the-import-of-cross-task-productivity.html)).

*Methodological caveat:* The Toner-Rodgers (2024) materials-discovery study (+44% novel materials) was publicly disavowed by MIT in May 2025 following data-integrity concerns. Widely cited but should not be treated as established fact.

## 5. Interaction Modes: Centaur, Cyborg, Self-Automator

Mollick's three-mode taxonomy is now empirically grounded:

**Centaurs** maintain clean human/AI role separation, handing off discrete tasks based on frontier mapping. **Cyborgs** intertwine human and AI continuously at sub-task granularity. Randazzo, Lifshitz et al. (HBS WP 26-036, 2026) added the **self-automator**: full delegation with periodic oversight. Empirical distribution across 244 BCG consultants: ~60% cyborg, ~30% centaur, ~10% self-automator.

Schoenegger, Park, Karger & Tetlock's superforecasting study (2024/2025, ACM TiiS, [link](https://eprints.lse.ac.uk/127059/3/3707649.pdf)) found **both well-calibrated and deliberately overconfident GPT assistants improved forecasting accuracy 23–43%** — suggesting much of the centaur gain comes from forced structured reasoning rather than AI advice quality. Combined with the historical chess record, this raises the question of **whether the centaur advantage is a transient regime** that disappears when AI exceeds humans on the full task, or a permanent feature of asymmetric cognitive strengths.

## 6. Automation Levels and Autonomy Frameworks

**Parasuraman, Sheridan & Wickens' (2000, [link](https://dl.acm.org/doi/10.1109/3468.844354)) four-function × ten-level model** remains the cleanest formal scaffolding. Their four automation functions — information acquisition, information analysis, decision/action selection, action implementation — map directly onto modern LLM workflow stages (RAG/retrieval, synthesis/analysis, recommendation, tool use/code execution). Yet no one has formally re-operationalized this for LLMs.

The 2023–2026 wave: **Morris et al.'s "Levels of AGI" (DeepMind, 2023, [arXiv:2311.02462](https://arxiv.org/html/2311.02462v2))** separates capability from autonomy across six levels. **Feng, McDonald & Zhang's "Levels of Autonomy for AI Agents" (2025, [arXiv:2506.12469](https://arxiv.org/abs/2506.12469))** defines five user-centered roles (Operator → Collaborator → Consultant → Approver → Observer) and is the most directly applicable to individual workflow design. Anthropic's "Measuring AI Agent Autonomy in Practice" (2025/2026, [link](https://www.anthropic.com/research/measuring-agent-autonomy)) surveys five competing frameworks empirically.

**Shneiderman's Human-Centered AI 2D framework** explicitly rejects the "more automation = less control" assumption — high automation and high human control can coexist (cameras, GPS, modern IDEs). This is a crucial conceptual move for knowledge work, where the goal is high-autonomy AI *with* high human oversight, not a trade-off between them.

The classical critique tempering all level-talk: **Dekker & Woods' "MABA-MABA or Abracadabra?" (2002)** — automation does not merely *replace* human work, it *transforms* it. The substitution myth is alive in current LLM discourse. Every sub-task offloaded to AI creates new monitoring, verification, and coordination work.

## 7. Cognitive Operation Taxonomies and Task-to-Tool Mapping

**Bloom's revised taxonomy** (Remember → Understand → Apply → Analyze → Evaluate → Create × factual/conceptual/procedural/metacognitive) is the most-imported cognitive framework in 2024–2026 LLM research. Empirically, **LLM capability decays sharply up the Bloom hierarchy**: BloomAPR (Ma et al. 2025, [arXiv:2509.25465](https://arxiv.org/pdf/2509.25465)) found ~81% success at Remember-level tasks, dropping to 43% at Apply and 13–41% at Analyze. Lee et al.'s CHI 2025 survey explicitly used Bloom's levels to show **GenAI shifts cognitive labor from lower-order production to higher-order verification, integration, and stewardship.**

**Cognitive Task Analysis** (CTA) methods (Crandall, Klein & Hoffman 2006 *Working Minds*; Militello & Hutton's ACTA, [PubMed](https://pubmed.ncbi.nlm.nih.gov/9819578/)) remain conspicuously underutilized. CTA is the canonical method for understanding what a knowledge worker actually *does* cognitively before allocating sub-tasks to AI — yet almost no production agent design uses it. Klein et al.'s macrocognition framework (sensemaking, problem detection, mental projection, coordination, [PMC](https://pmc.ncbi.nlm.nih.gov/articles/PMC5398123/)) is similarly absent despite obvious fit.

The cleanest bridge between classical cognitive architectures and LLM agents: Sumers, Yao, Narasimhan & Griffiths' **CoALA framework** (2024, TMLR, [arXiv:2309.02427](https://arxiv.org/abs/2309.02427)), mapping LLM agents onto modular memory (working/episodic/semantic/procedural), structured action spaces, and decision cycles drawn from ACT-R and SOAR.

**The practitioner world has a de facto task-routing approach that academia hasn't formalized.** Paterson's (2026, [link](https://ianlpaterson.com/blog/llm-benchmark-2026-38-actual-tasks-15-models-for-2-29/)) empirical benchmark of 15 models across 38 real daily tasks concluded that "routing beats model selection" — the generating function is a dispatch table matching task type to tool, not a single best model. This echoes Power's DSS taxonomy (model-/data-/knowledge-/document-/communication-driven systems) but is grounded in per-task empirical measurement rather than a priori categorization.

## 8. Practitioner Frameworks and Emerging Workflow Architectures

Practitioner literature is now driving the discipline. This section maps the most influential frameworks and where they converge.

**Mollick's centaur/cyborg/jagged frontier** ([link](https://www.oneusefulthing.org/p/centaurs-and-cyborgs-on-the-jagged)) and his book *Co-Intelligence* (2024) function as the dominant practitioner vocabulary. His four rules — always invite AI, be the human in the loop, give it a persona, assume this is the worst AI you'll ever use — are the closest to a widely adopted practitioner heuristic set.

**Karpathy's framework** (Software 1.0/2.0/3.0, jagged intelligence, anterograde amnesia, the autonomy slider, generator-verifier loop, [link](https://www.latent.space/p/s3)) gives precise vocabulary for coding workflows. The **autonomy slider** — instantiated in Cursor's Tab → Cmd+K → Agent Mode progression — is the clearest practitioner instantiation of what academic autonomy taxonomies describe abstractly: a per-action user control surface.

**Anthropic's agent-design patterns** ([link](https://www.anthropic.com/research/building-effective-agents)): prompt chaining, routing, parallelization, orchestrator-workers, evaluator-optimizer. Their multi-agent research system ([link](https://www.anthropic.com/engineering/multi-agent-research-system)) showed orchestrator-worker patterns outperformed single-agent Claude Opus by 90.2% at ~15× token cost. Their context-engineering guide ([link](https://www.anthropic.com/engineering/effective-context-engineering-for-ai-agents)) formalizes compaction, just-in-time retrieval, structured memory, and subagent isolation.

**Cognition's context-engineering principles** ([link](https://cognition.ai/blog/multi-agents-working)): the **read/write asymmetry** — multi-agent works for read-heavy tasks (research) but breaks for write-heavy tasks (code) unless writes are serialized. This is now consensus across Anthropic, LangChain, and Cognition.

**Coding-agent workflow patterns** have stabilized around three approaches: (1) continuous pairing (Cursor — taxes attention, preserves flow), (2) batch delegation (Devin — reduces presence, adds re-entry cost), and (3) spec-driven development (Harper Reed's spec → plan → execute loops, Amazon Kiro, GitHub Spec Kit). Claude Code's documented harness (CLAUDE.md memory, writer/reviewer, test-then-code) is the most comprehensive single-tool pattern.

**Personal knowledge management + AI** (Karpathy's "LLM Wiki," Obsidian + Claude Code patterns from Eric Ma and others) converges on: plain-text-as-substrate, persistent context files teaching personal taxonomy, reusable commands, inbox → process → integrate → review lifecycle. Every's Dan Shipper articulates this as the **"AI Sandwich"** (humans frame and review; AI does the middle) and **"Compound Engineering"** (plan → work → review → compound).

**Key practitioner insight lacking academic formalization:** Cowen's cross-task productivity bundling — per-task speedups don't translate proportionally to aggregate productivity because related tasks are productivity-linked. This connects directly to the Humlum & Vestergaard aggregate-zero puzzle.

## 9. Adjacent Fields: Imported, Underutilized, and Ripe for Bridging

**Joint Cognitive Systems / Cognitive Systems Engineering** (Hollnagel & Woods 2005) reframes human + LLM as a single coupled system. **Klein, Woods, Bradshaw, Hoffman & Feltovich's "Ten Challenges for Making Automation a Team Player" (2004, [PDF](https://www.jeffreymbradshaw.net/publications/17._Team_Players.pdf_1.pdf))** has become the most-cited pre-LLM paper in 2024–2026 agent design. Its requirements — Basic Compact, mutual models, predictability, directability, observability, goal negotiation, attention management, common-ground repair — function as the checklist for what an agent teammate needs. Xu & Gao's (2024, *Interactions*) **HAIJCS framework** is the cleanest bridge from CSE to LLM human-AI teaming.

**Distributed cognition** (Hutchins 1995) is being imported by Hutchins himself (Paris IAS 2024) and by Tao An's "Cognitive Workspace" (2025, [arXiv](https://arxiv.org/pdf/2508.13171)), which grounds LLM context management in Baddeley's working memory model. **The extended mind thesis** (Clark & Chalmers 1998) has been explicitly extended to LLMs by Smart, Clowes & Clark in *Synthese* 2025 ([link](https://link.springer.com/article/10.1007/s11229-025-05046-y)). Tong's 2026 survey ([arXiv](https://arxiv.org/pdf/2601.06030)) synthesizes the full Licklider–Engelbart–Clark lineage through to modern human-AI symbiosis.

**Engelbart's H-LAM/T** (1962) — Human using Language, Artifacts, Methodology, Training — is the most under-imported framework. It required co-evolution of all four components; current AI rollouts ship the artifact (the model) while methodology and training lag. Treating H-LAM/T as a literal rollout checklist would discipline most AI deployments.

**Other underutilized resources:** Power's DSS taxonomy for classifying AI tools by purpose; Nonaka's SECI cycle being extended to human-AI knowledge creation (Böhm & Durst 2025, Matsumoto et al.); Personal Information Management (Jones, Bergman & Whittaker) providing taxonomies for the "AI second brain" movement; **Endsley's situation awareness model** extended to human-AI teams in her own 2023 paper; and **Wickens' Multiple Resource Theory**, which would predict tool-stack attention overload (running Cursor + ChatGPT + a meeting simultaneously) but is absent from AI workflow research.

## 10. Key Researchers, Labs, and Thought Leaders

**Microsoft Research** has the deepest portfolio: Horvitz, Kamar, Amershi, Liao, Tankelevitch, Rintel, Sarkar, Bansal, Mozannar, Buçinca. The "Tools for Thought" research program ([link](https://www.microsoft.com/en-us/research/project/tools-for-thought/)) and the associated CHI 2025 workshop ([link](https://www.microsoft.com/en-us/research/blog/the-future-of-ai-in-knowledge-work-tools-for-thought-at-chi-2025/)) are the most concentrated effort on AI-augmented knowledge work. **Stanford HAI** covers empirical reliance work. **MIT CSAIL + Sloan/D³** drives both formal L2D theory (Sontag, Mozannar) and field experiments (Dell'Acqua, Lakhani). **Harvard SEAS/D³** hosts Buçinca, Gajos. **Wharton/HBS** bridges practice and research (Mollick, Lifshitz-Assaf, Kellogg). **CMU HCII**, **UW/AI2** (Weld, Fok), and **Stanford Digital Economy Lab** (Brynjolfsson) round out the empirical work.

**Adjacent-field bridgers:** Wei Xu (HAIJCS), Smart/Clowes/Clark (extended mind), Endsley (SA), Klein/Bradshaw/Hoffman/Feltovich (CSE/NDM), Tao An (cognitive workspace), Tong (augmentation→symbiosis).

**Practitioner thought leaders with formalized frameworks:** Mollick (Wharton), Karpathy (independent), Willison (independent — coined the canonical agent definition, [link](https://simonw.substack.com/p/i-think-agent-may-finally-have-a)), Schluntz & Zhang (Anthropic), Yan & Cognition Labs, Chase (LangChain), swyx (Latent Space, [link](https://www.swyx.io/about)), Shipper & Klaassen (Every, [link](https://every.to/chain-of-thought/compound-engineering-how-every-codes-with-agents)), Cowen (GMU, [link](https://marginalrevolution.com/)), and the Microsoft New Future of Work Report team.

## 11. Open Questions, Contested Ground, and Unfilled Gaps

**Stable consensus:** Explanations alone don't yield complementary performance. AI helps novices most on well-defined tasks. Cognitive forcing reduces overreliance with equity caveats. AI homogenizes outputs at the population level. Verification cost is the binding constraint. Workflow architecture predicts outcomes better than model choice.

**Genuinely contested:**

1. **Whether AI helps experts.** Skill-leveling (Brynjolfsson, Noy) vs. skill-amplifying (Otis high-baseline finding) vs. net-negative (METR). Likely resolution: the bottleneck differs — execution speed (AI levels skill) vs. judgment/filtering (AI amplifies whoever can already filter).
2. **Micro-to-macro translation.** 15–55% RCT gains coexisting with Humlum & Vestergaard's aggregate zero. Possible explanations: task reorganization absorbs time savings, cross-task bundling (Cowen), weak wage pass-through, measurement artifacts.
3. **Long-term cognitive effects.** Bastani's skill-atrophy evidence vs. Brynjolfsson's accelerated learning curves. The guardrail design matters more than the binary of AI access vs. none.
4. **Human+AI vs. AI alone in expert domains.** Goh's medical finding that AI alone wins vs. Everett's workflow-architecture fix. The claim appears workflow-dependent, not capability-dependent.
5. **Persistence of the centaur regime.** Chess history + Schoenegger's findings suggest centaur advantages may be transient as AI capability crosses task thresholds.

**What hasn't been formalized:**

- No normative framework for individual daily workflow choreography (when to consult, delegate, verify, refuse) — the Parasuraman (2000) equivalent for end users rather than system designers.
- Context engineering remains a practitioner discipline without academic theory.
- Multi-tool attention allocation lacks quantitative models despite Wickens' MRT being directly applicable.
- The interaction between agent autonomy level and metacognitive load is under-theorized.
- Long-term skill formation under continuous AI use lacks longitudinal data (most studies ≤6 months).
- Direct workflow-architecture comparison RCTs are rare; the field needs more studies structured like Everett 2025 and Bastani's guardrailed vs. unguardrailed designs.
- The feedback loop between task routing, skill development, and frontier migration over time (as you get better at using AI, the frontier shifts, changing optimal allocation) has no formal model.

## 12. Load-Bearing Assumptions and What Would Flip Them

Any formalization built on this landscape will inherit certain assumptions. Making them explicit now disciplines the next phase.

**Crux 1: "Workflow architecture > model capability."** This is the document's central claim. It's supported by Dell'Acqua (inside vs. outside frontier), Everett (workflow fix restoring physician+AI performance), and the general pattern that the same model yields very different outcomes under different interaction designs. But this claim is load-bearing on a specific regime: one where capability differences *between* frontier models are small relative to design differences *between* workflows. If a model capability jump is large enough that even naive workflows dramatically outperform expert workflows on current models, this claim inverts. *What would flip it:* A capability discontinuity (not incremental improvement) that eliminates the jagged frontier for a broad task class. The evidence base comes from a narrow window (2023–2025) of similar-capability frontier models — the claim may not survive a regime change.

**Crux 2: "The metacognitive bottleneck is the binding constraint."** This assumes production has been sufficiently automated that the bottleneck has shifted upward to planning, evaluation, and calibration. But for many knowledge workers, production *is* still the bottleneck — they lack the time, skill, or tool access to make AI-assisted production easy. The metacognitive framing may describe elite power users, not the median worker. *What would flip it:* Evidence that the majority of knowledge workers are production-constrained rather than metacognition-constrained, even with AI access. Lee et al.'s CHI 2025 survey (319 workers) partially supports the metacognitive framing, but the sample skews toward workers who already use AI regularly.

**Crux 3: "The jagged frontier is mappable and relatively stable."** Design principle #1 says "map your personal jagged frontier task-by-task." This assumes the frontier is stable enough to calibrate against. But if model capabilities shift every 3–6 months, the frontier migrates faster than a user can recalibrate. *What would flip it:* Evidence that frontier migration rate exceeds human calibration rate — that by the time you've learned where GPT-4 fails, GPT-5 has moved the boundary. The likely resolution is that the frontier has *stable topological features* (AI is reliably good at X-type tasks, reliably bad at Y-type) even as the boundary shifts, making the shape mappable even if the exact edge is volatile. This is an empirical question the field hasn't tested.

**Crux 4: "Verification cost is the binding constraint on appropriate reliance."** The rational-cost-benefit reframing (Vasconcelos, Fok & Weld) is load-bearing on approximate rationality — that people correctly estimate when verification is worth the effort. But if people are systematically miscalibrated about AI error rates (which the sycophancy finding from Randazzo et al. directly suggests), then the binding constraint isn't verification *cost* but verification *calibration*. The distinction matters for design: reducing cost helps a rational actor; improving calibration helps a miscalibrated one. Both interventions are different.

**Crux 5: "The individual knowledge worker is the right unit of analysis."** The entire document frames optimization at the individual level. But if the Humlum & Vestergaard aggregate-zero puzzle is explained by organizational dynamics (task reallocation, managerial absorption of time savings, coordination costs), then individual workflow optimization is locally optimal but globally insufficient. The right unit might be the team or the value chain. *What would flip it:* Evidence that individually optimized AI workflows produce organizational friction (e.g., faster individual output creating review bottlenecks downstream, or AI-homogenized outputs reducing team diversity of thought). The Anderson et al. (2024) homogenization finding and the organizational-absorption explanation for aggregate-zero both point in this direction.

**Crux 6: "The centaur/cyborg/self-automator taxonomy is durable."** It may instead be a transient artifact of current tool limitations. As tools evolve toward seamless human-AI blending (real-time co-editing, ambient AI, continuous context), the discrete modes may dissolve into a continuum. The taxonomy's value for formalization depends on whether the modes capture something structurally real about cognitive coupling or merely describe current interface affordances. *What would flip it:* Evidence that as tool integration deepens, the behavioral distinction between centaur and cyborg disappears — users naturally slide between modes within a single task rather than choosing one.

### Adversarial Challenge to the Project Framing

**Strongest objection:** "You're trying to formalize a workflow architecture for a system where one of the components (AI capability) changes faster than any formal model can track. By the time you've mapped the frontier, built the model, and tested it, the frontier has moved. The practitioner literature is ahead *precisely because* it doesn't try to formalize — it adapts via heuristics and rapid iteration. The academic aspiration to a 'unified normative framework' is a category error: this is an engineering problem requiring adaptive heuristics, not a science problem requiring formal models."

**Why this objection is partially right:** The objection correctly identifies that any model parameterized on a specific capability profile (GPT-4 is good at X, bad at Y) will go stale within months. Fixed allocation rules are doomed. The practitioner instinct to stay adaptive is sound.

**Why the strongest version of the project survives it:** Even in rapidly changing systems, structural invariants exist that a formal model *should* capture. The metacognitive bottleneck doesn't disappear when models improve — it shifts to new decisions. The verification cost trade-off doesn't change *shape* when models get better — the threshold moves. The automation ironies are structural properties of *any* delegation relationship between a principal and an imperfect agent. What a formal model should capture is the **generating function** — the invariant structure that produces the right allocation *given any capability profile* — not a specific allocation for a specific model. The model should be parameterized by capability, not dependent on a fixed capability level. This is exactly the difference between Parasuraman's (2000) framework (which has lasted 26 years despite massive automation changes) and any specific automation allocation table (which goes stale quickly). The target is a model that says "here is how to *decide* what to delegate" — not "here is *what* to delegate."

## 13. Design Principles Supported by the Current Evidence

The empirical and theoretical record converges on a set of actionable principles for designing an individual knowledge worker's AI workflow:

1. **Map your personal jagged frontier task-by-task.** Outside-frontier AI use is actively harmful, so the first design decision is calibrating which sub-tasks are inside and outside for *you specifically*. This frontier is personal (varies by expertise) and dynamic (shifts with practice and model updates).

2. **Match interaction mode to task structure.** Centaur (clean handoff) for tasks with verifiable checkpoints. Cyborg (interleaved) for creative or ill-structured work. "Independent-then-synthesize" for high-stakes expert judgment.

3. **Minimize verification cost, not maximize AI capability.** The binding constraint is verification, not generation. Design for structured outputs, confidence signals, and cheap-to-check formats.

4. **Insert deliberate friction at decision points.** Cognitive forcing functions (form your own view before seeing AI output) reduce overreliance. Sarkar's "Friction-Induced AI" concept shows this can be built into tool design.

5. **Treat context engineering as the central craft.** Practitioner consensus: the binding constraint is not model intelligence but what context the model operates in. CLAUDE.md files, system prompts, persistent memory, and structured instructions are higher-leverage than model selection.

6. **Route tasks, don't pick a single best tool.** Paterson's empirical result ("routing beats model selection") is the practitioner instantiation of L2D theory. Build a personal dispatch table matching task types to tools.

7. **Preserve skills with guardrails.** Hint-only AI in learning contexts. AI-free zones for capabilities you need to maintain. Bastani's guardrailed-AI design prevented skill atrophy while preserving performance gains.

8. **Serialize writes in agentic systems.** Cognition's read/write asymmetry: multi-agent is powerful for research/analysis but breaks for code/document production unless writes are serialized.

9. **Watch for the metacognitive bottleneck.** The limiting resource in AI-augmented work is no longer effort but judgment and attention allocation. Tankelevitch's framework suggests optimizing for metacognitive efficiency, not throughput.

10. **Budget for automation ironies.** Every sub-task delegated to AI creates new monitoring, verification, and coordination work. Simkute et al.'s four productivity-loss categories are predictable and designable-against.

---

## Model
*topic: technology-utilization-architecture · stage: model · pass 6 · in-progress*

Generator-verifier loop with a per-task autonomy slider. The per-task value function V(u, v; θ) decomposes into four orthogonal channels (quality, attention, risk, skill); V is exactly bilinear in (u, v) so per-task optima land at three corners — do-yourself, self-automator, spec-driven. Centaur and cyborg arise as aggregate-level patterns from cross-sub-task corner mixing. Portfolio aggregation under a daily attention budget surfaces the Lagrangian shadow price μ that reroutes longer tasks first when budget binds. Five cruxes, six Stage-4 fitting targets, three engaged objections. Interactive two-tab dashboard included.

## TLDR

The lit review documents a research landscape; the topology stripped it down to load-bearing structure. This stage formalises the cleanest target the topology surfaced: a **generator-verifier loop with a per-task autonomy slider**, designed to survive capability change rather than encode a snapshot of any specific model's frontier. The optimisation target is **output quality per unit of human attention** for an individual knowledge worker. The formalisation makes three moves at once — **decomposition** (four orthogonal value channels: quality, attention, risk, skill), **generating function** (a per-task value function whose corner solutions reproduce the five empirically observed workflow modes), and **integration** (a single formalism that composes Karpathy's slider, Mollick's typology, Vasconcelos verification-economics, Bastani's atrophy, Bainbridge's substitution myth, and Madras-Mozannar L2D into one object).

The per-task value function is `V(u, v; θ) = Q(u,v) − α·A(u,v) + λ·S(u,v) − σ·R(u,v)`, where `u` is the autonomy level (fraction of the task delegated to AI) and `v` is the verification depth (fraction of AI output independently checked). The four channels are conceptually distinct mechanisms: quality `Q` rewards letting the better agent do the work, with verified output achieving the complementary-product ceiling `c_⋆ = c_AI + (1 − c_AI)·c_H` (either AI was right or human catches the error); attention `A` charges for human time *and* for the irreducible monitoring cost even at full delegation (the L1 substitution-myth invariant baked in via `ε > 0`); risk `R` penalises uncaught AI errors at a rate proportional to stakes σ; skill `S` rewards practice and penalises unverified delegation (Bastani-style atrophy). V turns out to be *exactly bilinear* in `(u, v)` — collecting terms gives `V = K_0 + K_u·u + K_v·v + K_uv·u·v` — which means the maximum on the unit square is always at a corner. Three corners are candidates: `(0, 0)` do-yourself, `(1, 0)` self-automator, `(1, 1)` spec-driven (the corner `(0, 1)` is dominated since verifying with no AI involvement is pure cost). **Centaur and cyborg modes do not arise as per-task optima** — they appear only as aggregate-level patterns when a worker mixes corner policies across sub-tasks with heterogeneous θ. This is a substantive prediction, not a limitation.

The portfolio extension aggregates per-task decisions under a daily attention budget. The headline result the model is designed to produce is **S1** (workflow architecture > model capability): on the same task mix and same `c_AI`, optimal routing dominates "max-AI" (self-automate everything) and dominates the naive flat-cyborg heuristic (`u=0.7, v=0.3` everywhere). The bilinearity finding sharpens this: the naive flat-cyborg policy is exactly the failure mode — it applies an interior `(u, v)` value that the bilinear structure says no individual sub-task should land at. Optimal routing differentiates across tasks (different corners for different θ); the aggregate `(ū, v̄)` across the day looks interior because the corners differ, not because any single decision is interior. The generating function is **parameterised by capability** (the L3 invariant): if `c_AI` rises uniformly across tasks, the model rebalances; if it rises only on certain task types, the boundary of optimal `u*` shifts but the *shape* of the routing rule stays put.

This stage produces seven things: (1) a math object — `V(u, v; θ)` and the four-channel decomposition; (2) a workflow-mode classifier — three-corner per-task router plus a five-region label-partition of the (u, v) plane for observed worker behaviour; (3) a portfolio aggregator with budget-aware shadow-price routing (μ-binary-search over per-task `α_eff = α + μ·g`); (4) the interactive two-tab dashboard below; (5) **five cruxes** of the model (load-bearing claims whose collapse rebuilds it); (6) **six Stage-4 fitting targets** (parameter calibrations and qualitative predictions the data pipeline should test); (7) three engaged objections (`c_AI` unobservable, model just recovers practitioner intuitions, model is single-shot not strategic) with steelmen and what survives. Scope is explicit: the formalisation does *not* capture the aggregate-zero puzzle (E4/O2 — organisational dynamics), cross-task productivity bundling (G8/Cowen), `c_AI` miscalibration on novel tasks (O7), sycophancy as a verification-degrader (E13), or frontier migration over time (O4). These are named scope-limits, not silent assumptions.

<CognitivePartnershipModel client:load />

## How to read this stage

The dashboard above is the artifact. Everything below is the spec.

Two interactive surfaces:

1. **Per-task router.** Inputs: `(c_H, c_AI, φ, σ, λ)` for one task. Outputs: optimal `(u*, v*)`, the workflow-mode label, and a four-bar decomposition showing which channel dominates. Use this to answer "for a task with these characteristics, what's the right way to use AI?"

2. **Day portfolio.** Inputs: a basket of task types with counts. Outputs: total quality, total attention, total skill change under four strategies (always-self / max-AI / naive cyborg / optimal routing). Use this to see the S1 effect — same AI, different workflow architectures, very different outcomes.

If the dashboard says one thing and your gut says another, the diagnostic is to check (a) whether your `(c_H, c_AI)` estimates are calibrated and (b) whether the constants `(α, ε, β, M)` are calibrated for your task density. This is exactly what Stage 4 (the data pipeline) is for.

## 1. The formalisation moves

Three things this stage does — explicit so they're inspectable separately.

**Move 1 — Decomposition.** The per-task value `V` splits into four orthogonal channels (`Q`, `A`, `R`, `S`). "Orthogonal" here means: each channel responds to `(u, v)` in a distinguishable way, so when a slider moves, the user can see which channel is driving the change. This isn't a stylised choice — it's how the topology's mechanism nodes (G3, G7, G9, L1) actually compose. Without the decomposition, "the gain from AI" is a black box; with it, the user can ask "is this gain coming from quality, time-saved, or skill?" and answer.

**Move 2 — Generating function.** The five workflow modes (P1 / E10) are not given as a typology to be matched. They are *generated* as solutions to `argmax V(u, v; θ)` under different parameter regimes. This is the L3 invariant in operational form: change `θ`, the optimum moves, the mode label changes — but the *function that produces the mode* is fixed. A practitioner who hardcodes "use Cursor for boilerplate, ChatGPT for strategy" gets a lookup table; this gets a generating function.

**Move 3 — Integration.** Six prior objects compose into one: Karpathy's autonomy slider (P2 — the `u` axis), Shneiderman's 2D framework (L4 — the (autonomy, control) plane is the (u, v) plane), Vasconcelos verification-economics (G3 — the `−α·v·φ` term), Bastani's guardrail finding (E9 — the `−β·u·(1-v)` skill term), Bainbridge's substitution myth (L1 — the `ε > 0` residual attention), and Madras-Mozannar L2D (L2 — the joint-surface optimisation at portfolio level). None of these alone is the model; the model is what they jointly imply.

What's *not* yet ready for formalisation, kept in §9 as scope-limits: cross-task bundling (G8), organisational absorption (E4/O2), miscalibration of `c_AI` (O7), sycophancy as quality-degrader (E13), frontier migration over time (O4).

## 2. Variables and objects

**Decision variables (per task), continuous on the unit square:**

| Symbol | Range | Meaning |
|---|---|---|
| `u` | [0, 1] | Autonomy level — fraction of generation delegated to AI |
| `v` | [0, 1] | Verification depth — fraction of AI output independently checked |

**Task parameters** (the vector `θ`):

| Symbol | Range | Meaning |
|---|---|---|
| `c_H` | [0, 1] | Human capability on this task type |
| `c_AI` | [0, 1] | AI capability on this task type |
| `φ` | ≥ 0 | Verification-cost ratio (verify-time / generate-time) |
| `σ` | [0, 1] | Stakes — weight on uncaught-error penalty |
| `λ` | [0, 1] | Skill-formation value — how much the worker cares about preserving this skill |

**Constants** (calibrated, not per-task):

| Symbol | Default | Meaning |
|---|---|---|
| `α` | 1.00 | Attention price (normalisation) |
| `ε` | 0.15 | Residual attention at full delegation — L1 invariant (substitution myth) |
| `β` | 0.05 | Skill-atrophy rate per unit of unverified delegation |
| `M` | 0.08 | Per-task metacognitive routing tax — A1/G1 invariant |
| `c_⋆` | `c_AI + (1−c_AI)·c_H` | Verified-output ceiling — "either AI got it right OR human catches the error" (complementary product of independent error events) |

The constants are calibrated to the lit-review anchors (Mozannar CUPS for ε; Bastani 17% drop for β; Tankelevitch metacognitive load for M). They can be re-fit by Stage 4 against telemetry data.

## 3. The per-task value function

```
V(u, v; θ) = Q(u, v) − α·A(u, v) + λ·S(u, v) − σ·R(u, v)
```

### 3.1 Quality channel `Q`

```
Q(u, v) = (1 − u)·c_H + u·[(1 − v)·c_AI + v·c_⋆]
```

With probability `(1 − u)` the human did the generation and quality is `c_H`. With probability `u` the AI did the generation, of which fraction `(1 − v)` ships unverified at quality `c_AI` and fraction `v` is verified. Verified output achieves quality `c_⋆ = c_AI + (1 − c_AI)·c_H = 1 − (1 − c_H)·(1 − c_AI)` — the probability that *either* the AI got it right *or* (it didn't and) the human catches the error, treating the two error events as independent. Linear in `u` for fixed `v`; linear in `v` for fixed `u`.

The complementary-product form natively handles the deskilled-verifier limit (`c_H → 0` ⇒ `c_⋆ → c_AI` — verification adds nothing when the human can't recognise errors) and the verifier-stronger limit (`c_H → 1` ⇒ `c_⋆ → 1` — careful human verification approaches a quality ceiling). Pass 1 used `c_⋆ = max(c_H, c_AI)`, which over-credits verification when `c_H > c_AI` (as if the human catches every AI error) and under-credits it when `c_H < c_AI` (as if a partly-skilled human catches no AI errors). The form here treats both with one expression and one structural assumption (independence of human and AI error events).

Caveat: `c_H` enters both the generation cost (when `u = 0`) and the verification benefit (the extra catching power at `v > 0`). Karpathy's G9 generator-verifier-asymmetry says verification is typically *easier* than generation — recognising an error costs less than producing a correct answer from scratch. A more accurate model would carry a separate `c_V` (verifier capability) ≥ `c_H` for low-`c_H` workers. Held for Stage 4; named in §9 scope-limits.

### 3.2 Attention channel `A`

```
A(u, v) = (1 − u·(1 − ε)) + v·φ + M
```

Three pieces:

- `(1 − u·(1 − ε))`: human-side generation cost. At `u = 0`, the human does it all (cost = 1, the base generation time). At `u = 1`, residual attention `ε` remains — every offload creates monitoring/coordination work (Bainbridge L1). `ε > 0` is *the* L1 invariant in operational form: a model with `ε = 0` would predict that full delegation is free of attention cost, which is exactly the substitution myth.
- `v·φ`: verification cost. Linear in verification depth, scaled by the per-task verification-cost ratio `φ`. This is the G3 (Vasconcelos verification-economics) term.
- `M`: per-task metacognitive routing tax. Constant per task — classifying, choosing the workflow, monitoring AI for handoffs. Tankelevitch's metacognitive-demand finding (G1) compressed to a constant; Stage 4 can test whether `M` is task-type-dependent.

Note that ε is the *only* term that decouples attention from the (u, v) decision. It is what makes "max-AI" not free.

### 3.3 Risk channel `R`

```
R(u, v) = u·(1 − v)·(1 − c_AI)
```

Probability-of-uncaught-error: AI generated the output (`u`), it was not verified (`1 − v`), and the AI was wrong (`1 − c_AI`). Multiplied by stakes `σ` in the value function. This is the G2 (ironies-of-automation) term: rare critical errors get missed precisely when delegation is high and verification is low.

For high-σ tasks, this term is large enough to drive `v → 1` (the spec-driven / independent-then-synthesize regime, E8 — Everett 2025). For low-σ tasks, it's negligible and the optimum can sit at `v = 0` without harm.

### 3.4 Skill channel `S`

```
S(u, v) = (1 − u) − β·u·(1 − v)
```

Two terms:

- `(1 − u)`: practice — the human builds skill on the fraction of the work they did themselves.
- `−β·u·(1 − v)`: atrophy — unverified delegation erodes capacity. Verification preserves engagement (this is Bastani's "hint mode" finding E9: guardrails → no atrophy). The product `u·(1 − v)` is exactly the self-automator regime where atrophy is fastest.

`λ` is the worker's per-task valuation of preserving the skill. For tasks the worker explicitly wants to maintain capacity on (their core craft), `λ` is high and the skill term has bite. For boilerplate or one-off tasks, `λ` is low and skill is correctly ignored.

### 3.5 Putting it together

```
V(u, v; θ) =   (1 − u)·c_H + u·[(1 − v)·c_AI + v·c_⋆]            ← Q
             − α·[(1 − u·(1 − ε)) + v·φ + M]                       ← −α·A
             + λ·[(1 − u) − β·u·(1 − v)]                           ← +λ·S
             − σ·[u·(1 − v)·(1 − c_AI)]                            ← −σ·R
```

Five parameters in θ, two decisions, four channels. **Exactly bilinear** in `(u, v)`: collecting terms,

```
V(u, v; θ) = K_0 + K_u·u + K_v·v + K_uv·u·v
```

with

- `K_0 = c_H − α − α·M + λ`
- `K_u = (c_AI − c_H) + α·(1 − ε) − λ·(1 + β) − σ·(1 − c_AI)`
- `K_v = −α·φ` (K_v captures the cost of verification when there is no AI to verify, i.e., at `u = 0`; pure cost, hence ≤ 0 — this is exactly why corner `(0, 1)` is dominated by `(0, 0)` below)
- `K_uv = (c_⋆ − c_AI) + λ·β + σ·(1 − c_AI) = (1 − c_AI)·c_H + λ·β + σ·(1 − c_AI)` (always ≥ 0 — verification gain is monotone in `u`)

Bilinear functions on a unit square attain their maximum at a corner. The interior critical point, when it exists, is a saddle (Hessian eigenvalues `±K_uv`). So the per-task optimum is at one of the four corners — and since `K_v < 0`, `(0, 1)` is dominated by `(0, 0)` (verifying when no AI is involved is pure cost). The three meaningful corners:

- **`(0, 0)`** — do-yourself.
- **`(1, 0)`** — full delegation, no verification (self-automator).
- **`(1, 1)`** — full delegation with full verification (spec-driven / independent-then-synthesize).

Which corner wins depends on the signs of `K_u`, `K_u + K_uv + K_v`, and `K_v + K_uv` — three linear comparisons over θ.

## 4. Optimal policy: the three corners that win

V is bilinear → max at a corner. Of the four corners, `(0, 1)` is dominated by `(0, 0)` because `K_v < 0` (verifying with no AI involvement is pure cost). Three candidates remain — and a clean decision tree determines the winner:

- **Spec-driven `(1, 1)`** wins iff `K_u + K_v + K_uv > 0` *and* `K_v + K_uv > 0`.
- Else **self-automator `(1, 0)`** wins iff `K_u > 0`.
- Else **do-yourself `(0, 0)`** wins.

Equivalently: spec-driven dominates self-automator when `α·φ < K_uv` (verification cost is below benefit at full delegation: `α·φ < (1 − c_AI)·c_H + λ·β + σ·(1 − c_AI)`); self-automator dominates do-yourself when `K_u > 0` (the attention savings + AI quality gain outweigh skill loss + stakes risk).

**Comparative statics — what moves the corner choice:**

| Parameter increase | Effect on `u*` | Effect on `v*` (given `u* = 1`) |
|---|---|---|
| `c_AI − c_H` ↑ via `c_AI` rising (`c_H` fixed) | ↑ (K_u rises) | ↓ (K_uv falls: less verification benefit) |
| `φ` ↑ (verification expensive) | weakly ↓ via v\*-flip | ↓ (K_v more negative) |
| `σ` ↑ (stakes) | weakly ↓ (K_u falls by `σ·(1−c_AI)`) | ↑ (K_uv rises by `σ·(1−c_AI)`) |
| `λ` ↑ (skill matters) | weakly ↓ (K_u falls by `λ·(1+β)`) | ↑ slightly (K_uv rises by `λ·β`) |
| `c_AI` ↑ alone | ↑ | ↓ (both `(1−c_AI)·c_H` and `σ·(1−c_AI)` fall) |

This is the L3 invariant in tabular form: if `c_AI` rises uniformly, optimal `u*` rises and `v*` falls. The structural shape of the rule does not change. Two signs worth flagging because they're non-obvious: more reliable AI (`c_AI` ↑) *reduces* verification benefit (fewer errors to catch); higher stakes (`σ` ↑) pushes `u*` *down* (refuse AI when stakes are high) *and* `v*` *up* (if you do use AI, verify carefully) — a real tension the bilinear structure makes explicit.

### The five practitioner modes — labels for the (u, v) plane

The practitioner literature names five workflow modes. They span the (u, v) plane and are useful as **vocabulary for labelling observed worker behaviour** at any `(u, v)`:

| Region | Mode | Practitioner anchor |
|---|---|---|
| `u ≈ 0` | **Do-yourself** | (no-AI / refuse-AI) |
| `u ∈ (0, 1)`, `v` high | **Centaur** | Mollick — clean handoff with verification gate |
| `u ∈ (0, 1)`, `v` low/mid | **Cyborg** | Mollick — interleaved, partial verification |
| `u ≈ 1`, `v` high | **Spec-driven / independent-then-synthesize** | Everett 2025; Compound Engineering |
| `u ≈ 1`, `v` low | **Self-automator** | Randazzo HBS 26-036 (the trap) |

The per-task router only returns the three corners — `(0, 0)`, `(1, 0)`, `(1, 1)` — corresponding to do-yourself, self-automator, and spec-driven. **Centaur and cyborg do not arise as per-task optima.** They appear empirically as **aggregate-level labels** when a worker mixes corner policies across sub-tasks with heterogeneous θ — some sub-tasks done alone, others fully delegated; some AI outputs verified, others shipped. The day-level `(ū, v̄)` averages out to interior values that get labelled "cyborg" or "centaur" depending on the ratio. The Day Portfolio tab demonstrates this directly.

This is a substantive prediction of the formalisation, not a limitation. Bilinearity is faithful to the structure of the problem: V collects to one bilinear form `V = K_0 + K_u·u + K_v·v + K_uv·u·v`, even though three of the four channels (Q, R, S) carry their own `u·v` terms — they sum to one consolidated `K_uv`. The model says: at any single sub-task with a single θ, pick a corner — fully delegate or don't, fully verify or don't. The naive flat-cyborg strategy (`u = 0.7, v = 0.3` applied to every task) is exactly the failure mode — applying an interior policy uniformly is what no individual sub-task should do under bilinearity.

## 5. Portfolio aggregation — the day

A worker faces N tasks per day, each with its own `θ_i`. Total attention budget `T`. The portfolio problem:

```
maximise   Σ_i Q_i(u_i, v_i) · count_i
subject to Σ_i A_i(u_i, v_i) · g_i · count_i ≤ T
```

where `g_i` is the task type's base generation time. The Lagrangian gives a shadow price `μ ≥ 0` on the budget constraint, and the per-task choice rule becomes

```
argmax V_i − μ·A_i·g_i  =  argmax V_i with α replaced by α_eff_i = α + μ·g_i
```

— that is, **budget pressure raises the effective attention price, more so for longer-base-time tasks**. When `μ = 0` the budget is slack and per-task choice is unconstrained `argmax V`; when `μ > 0` the budget binds and `α_eff` rises until total absolute attention `Σ A_i·g_i·count_i` fits `T`. The biasing is structural: raising `α_eff` raises `K_u` (self-automator becomes more attractive vs. do-yourself) and makes `K_v = −α_eff·φ` more negative (verification becomes less attractive). So as the budget tightens, longer tasks reroute first from spec-driven `(1, 1)` to self-automator `(1, 0)` — the lowest-A corner.

This is the L2 invariant operationalised at portfolio level. The naive practitioner rule "use AI when AI is better" compares `c_H` to `c_AI` task-by-task in isolation; the joint-surface rule compares marginal `Q` per unit attention saved against the day's shadow price. They give the same answer when attention is abundant (`μ ≈ 0`); they diverge sharply when the day is tight.

**Strategies the dashboard compares:**

1. **Always-self.** `u_i = 0` for all i. No AI, no atrophy, no verification cost — but no productivity gain. The pre-AI baseline.
2. **Max-AI.** `u_i = 1, v_i = 0` for all i. The corner uniformly applied. Fast, but high risk on high-σ tasks and accelerated atrophy on high-λ tasks.
3. **Naive cyborg.** `u_i = 0.7, v_i = 0.3` for all i. A flat *interior* policy applied uniformly — exactly what the bilinear structure says no individual task should land at. Interior values arise legitimately only as averages over heterogeneous-θ sub-tasks. Applying them uniformly violates the structure and underperforms.
4. **Optimal routing (budget-aware).** Per-task corner choice under the shadow-price-adjusted `α_eff_i = α + μ·g_i`, with `μ` solved by binary search until the budget is met (or until even uniform self-automator overflows). Each task lands at the corner appropriate to its θ *and* the day's binding budget. The aggregate `(ū, v̄)` across the day looks interior because different tasks land at different corners; the dashboard surfaces `μ` below the strategy table so the user can see when the budget is biting.

The headline prediction (S1): **optimal routing dominates max-AI by quality-per-attention and dominates always-self by attention efficiency, on the same `c_AI`**. The gap between optimal routing on mid-tier AI and naive-flat routing on frontier AI is the empirical bound the model places on "workflow architecture > model capability." Mechanism: optimal routing differentiates across tasks (different corners for different θ) and reroutes under budget pressure (shadow price `μ`); naive uniformly applies an interior policy that is structurally never the per-task optimum at any α.

## 6. Calibration anchors

Where the constants come from. None of these is precise; all are Stage-4 fitting targets.

- **`α = 1`** — normalisation. Attention is the numéraire; all other costs are denominated in attention units.
- **`ε = 0.15`** — Mozannar CUPS (E12) shows verification + monitoring is a "substantial fraction" of total interaction time even when AI is doing the generation. 15% is a midpoint of the reported range; Stage 4 should tighten this from telemetry.
- **`β = 0.05`** — Bastani's 17% unassisted-performance drop (E9) over a session of ~30 unguardrailed tasks → ~0.5% atrophy per task at `u = 1, v = 0`. Setting `β = 0.05` means the model implies ~5% atrophy per task in the worst-case regime, which compounds to Bastani's order-of-magnitude over a week. The lit review explicitly notes most studies are under 12 months; β at the per-task scale is what compounds to the longitudinal scale.
- **`M = 0.08`** — Tankelevitch (G1) finds metacognitive load is the binding constraint for AI users; CUPS (E12) finds verification + planning consume a meaningful fraction of total time. 8% per task is a calibration anchor consistent with the magnitude of the metacognitive-bottleneck claim. Stage 4 should test whether `M` varies by task type (it likely does — high-stakes strategic decisions have larger M than routine email).

At the per-task level, the defaults predict `(1, 0)` self-automator at routine-low-stakes corners (Randazzo's ~10% empirically), `(1, 1)` spec-driven where verification benefit dominates verification cost, and `(0, 0)` do-yourself in outside-frontier regimes. The full Randazzo 60/30/10 (cyborg/centaur/self-automator) distribution emerges only at the *aggregate* level, when a worker's day mixes corner policies across heterogeneous-θ sub-tasks — see §4. The defaults predict outside-frontier harm (Dell'Acqua E3) when `c_H > c_AI` and the worker mis-routes to `u > 0` (the model's prescription is `u* = 0` there); the harm is from disobeying the optimum, not from a model output.

## 7. Worked anchors against the empirical record

Six probes. Each picks a parameter regime, computes the corner optimum exactly, and compares to the lit-review anchor.

**A. Brynjolfsson (E1, E2) — customer-service novice +34%, top performers ~0%.**

*Novice:* `θ = (c_H = 0.40, c_AI = 0.70, φ = 0.20, σ = 0.20, λ = 0.10)`. `c_⋆ = 0.70 + 0.30·0.40 = 0.82`. Then `K_u = 0.30 + 0.85 − 0.105 − 0.06 = 0.985 > 0` (full delegation beats do-yourself), and `K_v + K_uv = −0.20 + (0.12 + 0.005 + 0.06) = −0.015 < 0` (verification cost barely exceeds benefit). **Optimum: `(1, 0)` self-automator.** Quality lift over always-self: `c_AI − c_H = +0.30`. Direction matches the +34% Brynjolfsson finding (which is in resolution rate, mixing speed and accuracy).

*Expert:* `θ = (c_H = 0.85, c_AI = 0.70, φ = 0.20, σ = 0.20, λ = 0.10)`. `c_⋆ = 0.70 + 0.30·0.85 = 0.955`. Then `K_u = −0.15 + 0.85 − 0.105 − 0.06 = 0.535 > 0`, and `K_v + K_uv = −0.20 + (0.255 + 0.005 + 0.06) = +0.12 > 0` (verification benefit dominates because `c_⋆ − c_AI = 0.255` is large — a skilled human catches AI errors). **Optimum: `(1, 1)` spec-driven.** Quality lift: `c_⋆ − c_H = +0.105`, only +12% relative.

So the model produces: novice at full-delegate-no-verify (Q = 0.70 from c_AI alone), expert at full-delegate-full-verify (Q = 0.955 = AI augmented by skilled human catching). Both delegate; the difference is in verification depth. The empirical "expert ~0%" finding reflects throughput ceiling (experts already at maximum call rate, can't redeploy saved attention to more calls) rather than zero quality lift — a context the model doesn't carry.

**B. Dell'Acqua BCG (E3) — outside-frontier 19-pp quality drop.** `θ = (c_H = 0.70, c_AI = 0.40, φ = 0.30, σ = 0.50, λ = 0.50)`. `c_⋆ = 0.40 + 0.60·0.70 = 0.82`. Then `K_u = −0.30 + 0.85 − 0.525 − 0.30 = −0.275 < 0`. **Optimum: `(0, 0)` do-yourself.** If the worker mis-routes to `(1, 0)`, quality drops from `c_H = 0.70` to `c_AI = 0.40` — a 30 pp loss. The empirical 19 pp reflects partial mis-routing (some subjects partially used AI, some didn't); the model's prediction is a clean upper bound on the harm.

**C. Bastani PNAS (E9) — guardrails preserve skill.** Generic learning task with high skill-formation: `θ = (c_H = 0.40, c_AI = 0.70, φ = 0.30, σ = 0.20, λ = 0.50)`. `c_⋆ = 0.82`. `K_u = 0.30 + 0.85 − 0.525 − 0.06 = 0.565 > 0`. `K_v + K_uv = −0.30 + (0.12 + 0.025 + 0.06) = −0.095 < 0`. **Optimum: `(1, 0)` self-automator.**

This is a real and honest finding: at lit-review-anchored constants (β = 0.05), the skill-preservation push toward verification (`λ·β·u = 0.025` at u = 1) is too small to flip the corner against verification cost. The guardrail effect is *structurally present* — at the (1, 1) corner, `S = 0` (no atrophy) versus `S = −0.05` at (1, 0), so `λ·ΔS = +0.025` — but the verification cost (`α·φ = 0.30`) dominates. Numerically, self-automator beats spec-driven by `α·φ − K_uv = 0.30 − 0.205 = 0.095` net at default constants. To make guardrails decisive (flip the corner to spec-driven), the model needs `λ·β > α·φ − [(1 − c_AI)·c_H + σ·(1 − c_AI)] = 0.30 − 0.18 = 0.12`. At default `λ = 0.50`, `β` must exceed `0.24`; at `β = 0.05`, `λ` alone cannot flip the corner (would require `λ > 2.4`, impossible since `λ ∈ [0, 1]`); at default `β = 0.05` and `λ = 0.50`, lowering `φ` flips the corner once `φ < 0.205` — barely cheaper verification than the default 0.30.

**Honest reading:** the model says the guardrail effect is real but weak at default constants. Stage-4 fitting target Q2 is whether `β` should be larger to match Bastani's empirical magnitude. This is not a model failure — it's the model surfacing a calibration question the lit review left implicit.

**D. Everett (E8) — independent-then-synthesize restores complementarity.** `θ = (c_H = 0.65, c_AI = 0.70, φ = 0.40, σ = 0.90, λ = 0.30)`. `c_⋆ = 0.70 + 0.30·0.65 = 0.895`. `K_u = 0.05 + 0.85 − 0.315 − 0.27 = 0.315 > 0`. `K_v + K_uv = −0.40 + (0.195 + 0.015 + 0.27) = +0.08 > 0`. **Optimum: `(1, 1)` spec-driven.** Mechanism: the `σ·(1 − c_AI) = 0.27` term in `K_uv` makes verification valuable precisely because stakes are high and AI is fallible. Matches Everett's lit-review story exactly.

**E. Randazzo self-automator (E10) — ~10% of consultants in the trap.** Routine consulting task: `θ = (c_H = 0.70, c_AI = 0.85, φ = 0.20, σ = 0.20, λ = 0.10)`. `c_⋆ = 0.955`. `K_u = 0.15 + 0.85 − 0.105 − 0.03 = 0.865 > 0`. `K_v + K_uv = −0.20 + (0.105 + 0.005 + 0.03) = −0.06 < 0`. **Optimum: `(1, 0)` self-automator.** Self-automator is the correct policy at this θ — the empirical finding "~10% of consultants are self-automators" is about θ-distribution (~10% of work-instances have these characteristics), not about systematic mis-routing. A misread of Randazzo's data as "self-automator is always wrong" is a category error the model corrects.

**F. Schoenegger (E18) — even overconfident GPT improves forecasting +23–43%.** This is *outside* the model's current formalism. The model attributes any gain at `u > 0` to `c_AI` (advice quality), but Schoenegger's finding suggests structured reasoning is doing significant work independent of the AI's confidence calibration. A constant `+δ` to Q whenever `u > 0` would represent this — held for future passes if it changes downstream predictions. **Honest gap.**

### Cross-context note: outcome heterogeneity across these anchors

The six papers above measure different outcome variables. The model's `Q` ("probability of correct/high-quality output") maps cleanly onto Dell'Acqua, Everett, and Schoenegger (all output-quality measures). Brynjolfsson's "issues resolved per hour" maps approximately onto `Q × call-rate`, where call-rate depends on attention saved (`A`); the model's `+34% novice` prediction is a Q-only claim, while Brynjolfsson's empirical 34% bundles throughput. Bastani's "unassisted retest performance" is a downstream effect of the `S` channel accumulated over many task instances, not a per-task `Q` measurement. Randazzo's "behavioural-mode distribution" is a categorical prediction over which corner the worker chooses, not a quality measure at all. Mozannar's CUPS data is process telemetry (time fractions across interaction states), informative for `α` and `ε` calibration but not for `Q`.

Stage 4 must disaggregate which constants get fit against which outcome types — pooling them as a single calibration target would silently average over methodological apples and oranges. This is named explicitly as Q1–Q6 in §11.

## 8. The five cruxes

Load-bearing claims of the model. Collapse rebuilds it.

**C1 — Two-axis decision space (u, v).** The decision is reduced to autonomy and verification depth. If the actual workflow choice space has more dimensions that matter — e.g., context-engineering depth, prompt-iteration count, tool selection — the model is incomplete. *What would flip it:* empirical evidence that two workflows with identical `(u, v)` but different context-engineering produce systematically different outcomes (which the practitioner literature suggests is real — S4).

**C2 — Verification effectiveness equals generation skill.** `c_⋆ = c_AI + (1 − c_AI)·c_H` treats `c_H` as both generation skill *and* verifier capability — a worker who'd produce 40%-quality output alone catches 40% of AI errors. Karpathy's G9 generator-verifier-asymmetry says verification is typically *easier* than generation; a separate `c_V ≥ c_H` parameter would make low-`c_H` workers more effective verifiers. *What would flip it:* empirical evidence that verifier-recognition rates are uncorrelated with generation skill (would require introducing `c_V`). Deskilled-verifier worry is partially handled by the formula already (`c_H → 0` ⇒ `c_⋆ → c_AI`), but the *mechanism* of skilled-but-not-creative verifier (the editor archetype) is not in scope.

**C3 — `ε > 0` is the right operationalisation of L1.** The substitution-myth invariant is captured as residual attention at full delegation. If the actual structure is more like "delegation creates new tasks of comparable effort" rather than "delegation creates monitoring overhead", a constant ε is wrong shape. *What would flip it:* evidence that delegation-induced work scales with task complexity, not as a flat constant.

**C4 — Skill atrophy `β` is task-type-uniform.** All tasks atrophy at the same rate per unit of `u·(1-v)`. Some skills (e.g., motor-procedural) atrophy slower than others (verbal-fluency, calibration). *What would flip it:* longitudinal data on skill-specific atrophy rates under controlled AI-use exposure.

**C5 — Tasks are independent in the portfolio.** `Σ_i V_i` aggregates linearly. Cross-task productivity bundling (G8/Cowen) violates this — productivity gains on related tasks are correlated, not additive. *What would flip it:* empirical evidence that observed aggregate productivity (E4 / Humlum-Vestergaard zero) is driven by task-coupling effects, not just by individual mis-routing. This is the most likely-to-flip crux: the aggregate-zero puzzle is a smoking gun for it.

## 9. Scope limits — what the model does NOT capture

Honest disclosure of where the formalisation stops.

- **Aggregate-zero puzzle (E4 / O2).** Humlum-Vestergaard's zero across 25,000 Danish workers is *organisational*, not individual — task reorganisation, managerial absorption, coordination costs. The model is individual-level (the A6 crux of the topology); it cannot rebut or explain E4 directly. This is a sibling-artifact problem (organisational-level model), not a parameter to set. The model is locally optimal at the individual level; it is silent about whether aggregate effects emerge.
- **Cross-task bundling (G8).** `V` is summed across tasks; in reality `V_i` and `V_j` covary when the tasks are productivity-linked. C5 names this; the model does not encode it.
- **Calibration error on `c_AI` (O7).** `c_AI` is treated as known. In practice users are systematically miscalibrated (E16 — higher AI confidence → less critical thinking; E13 — sycophancy escalation). The model's `c_AI` should be the user's *belief* about AI capability; if belief and reality diverge, the model is locally optimal under a wrong belief. The topology's O7 (verification cost vs. verification calibration) is the natural extension.
- **Sycophancy as quality-degrader (E13).** The model assumes verification only helps. Randazzo HBS 26-021 documents AI flipping correct human judgments under pushback — verification can *worsen* outcomes when sycophancy escalates. Not encoded; would require a `c_⋆` that depends on the human's resistance to AI-pushback.
- **Frontier migration (O4).** `c_AI` is static within a session. Over months the frontier moves; the user must recalibrate. The model is a snapshot — extending it to a dynamic version would couple `c_AI(t)` to a learning model of the user's frontier-mapping rate.
- **Multi-tool attention interference (G10 — Wickens MRT).** The model is per-task; it does not capture the cost of running Cursor + ChatGPT + Slack simultaneously. The topology added G10 specifically to flag this; Stage 4 / Stage 5 should test whether a `M_concurrent(N_tools)` extension is needed.
- **Verifier skill ≠ generation skill (Karpathy G9).** The model uses `c_H` as both generation capability and verification effectiveness. Empirically, verification is often easier than generation — recognising an error costs less than producing a correct answer. A `c_V ≥ c_H` parameter would let low-`c_H` workers benefit from verification more than the current formula predicts. Held for Stage 4 / future passes; the C2 crux names this.
- **Partial verification ("skim") rounds to full or none.** Bilinearity makes `v* ∈ {0, 1}`. The empirical reality of skimming (partial-depth verification at reduced cost and reduced effectiveness) does not map onto the model — the model rounds skim up to full-verify when verification is cheap and down to no-verify when it's expensive. A future variant with convex verification cost (`v²·φ` instead of `v·φ`) or with a separate verification-depth-vs-effectiveness curve would produce interior `v*` solutions. Not added in this pass to preserve bilinearity's analytical clarity.

These eight are not bugs of the formalisation. They are the boundary of what an individual-task bilinear generator-verifier loop can carry. Beyond it lies organisational design, dynamic learning, and team-level cognition — sibling artifacts.

## 10. Adversarial + steelman

Three objections the formalisation has not yet engaged head-on, and the strongest version that survives each.

### Objection 1: `c_AI` is unobservable, so the model is unactionable on novel tasks.

**Steelman.** The optimal policy depends on `c_AI`, but in any new task or unfamiliar domain the worker doesn't know `c_AI` ahead of time. Calibrating `c_AI` requires running the task and verifying the output — but the model says "compute `argmax V` using `c_AI` you don't have." For experienced task types `c_AI` is calibratable from history; for genuinely novel tasks (much of knowledge work) it isn't. The Vasconcelos / Fok & Weld verification-economics frame already captures this — engagement is rational when verification is cheap; verifying a single AI output IS your way of measuring `c_AI` for that task type. The model assumes the calibration question is solved when it's actually the binding constraint.

**Why partially right.** For genuinely novel tasks where `c_AI` is unknown, the model cannot make a precise recommendation. Stage-4 fitting target Q3 (mode-distribution match against Randazzo BCG) implicitly assumes calibrated `c_AI` distributions across BCG-like tasks — only valid for well-studied domains.

**Why the strongest version survives.** The model doesn't need precise `c_AI`; it needs robustness across `c_AI` ranges, and the corner structure is robust. For almost any `c_AI < 0.4` in a high-stakes regime (`σ ≥ 0.7`), the corner is do-yourself; for almost any `c_AI > 0.8` in low-stakes routine (`σ ≤ 0.2`, low `λ`), the corner is self-automator. The "interesting" boundary regions — where `c_AI` uncertainty matters — are precisely the spec-driven regions. So the model's prescription under uncertainty is: **set `v = 1` to verify and learn**. The spec-driven corner doubles as a Bayesian-update mechanism; the verification cost `α·v·φ` is the explicit price of resolving the uncertainty. Stage 4 should formalise the explore-vs-exploit dynamics this implies (Q6 in §11).

### Objection 2: The model recovers practitioner intuitions and adds no new predictions.

**Steelman.** Mollick, Karpathy, Anthropic, and Cognition collectively say "match the workflow to the task." The model's predictions of corner solutions match Randazzo's empirical mode distribution. So what is the formalism contributing beyond a formal scaffold for intuitions practitioners already had?

**Why partially right.** Many model predictions match practitioner consensus on headline conclusions. Self-automator for routine, do-yourself for novel-and-high-stakes, spec-driven for high-stakes-with-cheap-verify — practitioners already say these.

**Why the strongest version survives.** The model contributes five things the practitioner literature does not:

1. **Quantitative trade-offs.** The model says how much worse self-automator is than spec-driven at default constants — at the Bastani anchor specifically, `0.095` net (`α·φ − K_uv = 0.30 − 0.205`). Practitioner literature is qualitative; this is parametrically calibratable, and Stage 4 will pin the constants.
2. **Non-obvious simultaneous constraints.** Higher stakes (`σ ↑`) push BOTH `u` DOWN (refuse AI for high-stakes work) AND `v` UP (if you do use AI, verify carefully) — a simultaneous prescription the practitioner literature does not make explicit. §4's comparative-statics surfaces this; intuition often conflates the two.
3. **Naive cyborg as structural failure mode.** Bilinearity says applying interior `(u, v)` uniformly is what no individual sub-task should do. This is a sharper criticism than "be thoughtful about which mode you use" — it identifies a specific failure mode (the BCG cyborg majority running flat-`(0.7, 0.3)` policies) and says they're structurally wrong, not just suboptimal.
4. **Budget-aware shadow-price reformulation.** The `μ` mechanism says: under attention scarcity, reroute longer-base-time tasks first (since `α_eff_i = α + μ·g_i` rises proportionally with `g_i`). Practitioner literature has nothing like this prescriptive structure for portfolio-level decisions.
5. **Stage-4 calibration targets.** The model identifies six specific empirical questions (Q1–Q6 in §11) that fit a generator-verifier dispatch framework. Practitioner heuristics are unfalsifiable by design — they update without preserving the reasoning, so users can't tell when a heuristic stops applying. The L3 invariant is the antidote.

The model recovers practitioner intuitions on top-line conclusions and adds quantitative, edge-case, and falsifiable structure beyond. The contribution is in calibration, simultaneous constraints, structural failure modes, portfolio-level shadow pricing, and Stage-4 testability — not in inventing new top-level recommendations.

### Objection 3: The model is single-shot; the topic question is dynamic.

**Steelman.** "Optimal configuration for an individual knowledge worker" is implicitly a static answer to a fundamentally dynamic question — AI capability shifts month-over-month, the worker's skill atrophies under sustained delegation, calibration on `c_AI` drifts as new model versions ship. A static dispatcher with parametric flexibility doesn't tell the worker how to *anticipate* and *prepare for* capability change.

**Survives because** (a) Parasuraman, Sheridan & Wickens' (2000) function-allocation framework has been useful for 25+ years despite being static — parametric statics IS what's wanted from a generating function (the L3 invariant); (b) the trajectory questions properly belong to sibling topics in the LLM Iterate roster — `navigating-ai-world` for AI-induced trajectory of skill/meaning/relational channels, the planned `prediction-calibration` topic for `c_AI` calibration drift, the planned `bedrock-generating-functions` for the temporal-aggregation patterns this and other models share. Static-but-parameterised is the right scope here; dynamic extensions are cross-topic by design, not gaps in the present formalisation.

## 11. Stage-4 fitting targets

Six named questions Stage 4 should test against data.

**Q1 — `(α, ε)` from CUPS telemetry.** Mozannar CUPS gives time fractions across coding interaction states. Calibrate `ε` from observed monitoring-time-at-high-`u`; calibrate `α` from observed cyborg-vs-centaur time efficiency.

**Q2 — `β` from Bastani longitudinal regime.** Bastani's 17% drop is one window. Longer-window data (Lee-Sarkar CHI 2025; Anti-Social Century) should pin per-task atrophy rate. The model predicts: `β` should be ~10× larger for skills the worker uses heavily than for tasks they delegate occasionally.

**Q3 — Mode-distribution match against Randazzo BCG.** Given a θ-distribution prior over BCG-consultant tasks, does the optimal-routing distribution over (u\*, v\*) match the observed (~60% cyborg / ~30% centaur / ~10% self-automator)? If not, which constant is mis-fit?

**Q4 — Outside-frontier harm prediction (Dell'Acqua replication).** For subjects forced to use AI on `c_H > c_AI` tasks, the model predicts a quality drop proportional to `u·(c_H − c_AI)`. Stage 4 should test the linearity and the slope.

**Q5 — Workflow-architecture-vs-capability bound.** The headline S1 prediction. Construct two simulated workforces: (a) frontier AI + naive flat-cyborg routing, (b) mid-tier AI + optimal routing on the same task mix. The model predicts (b) outperforms (a) on net quality once `c_AI_naive` falls below some threshold. Stage 4 should locate the threshold and test against Everett 2025 / Dell'Acqua at that precision.

**Q6 — Calibration uncertainty and explore-vs-exploit on `c_AI`.** The model treats `c_AI` as a known input. In practice, workers learn `c_AI` by running the task and verifying the output (Vasconcelos verification-economics). For novel tasks, the spec-driven corner doubles as a calibration mechanism — verification reveals `c_AI` for next time. Stage 4 should formalise the explore-vs-exploit structure: what is the optimal exploration premium when `c_AI` variance is high? When does verifying-to-learn dominate verifying-to-quality-control? The §9 scope-limit on `c_AI` miscalibration becomes a structural extension via this question, not just an acknowledged gap. Engages the §10 Objection 1 directly.

### Data starting points for each Q

The natural starting point for each fitting target is the lit-review paper(s) that motivated the corresponding parameter. Honest notes on likely data availability:

- **Q1 (`α`, `ε`)** — Mozannar CUPS 2024 (E12) telemetry across coding interaction states. Process-level data is likely Microsoft Research-internal; replication via Cursor / Claude Code anonymized usage logs is the alternative path.
- **Q2 (`β`)** — Bastani PNAS 2025 (E9) is the primary anchor; PNAS supplementary materials likely include the longitudinal panel needed to estimate per-task atrophy. Lee-Sarkar CHI 2025 (E16) provides multi-task panel context for a complementary fit.
- **Q3 (mode distribution)** — Randazzo HBS WP 26-036 (E10) for the 60/30/10 BCG distribution. HBS supplementary materials may include individual-level mode tags; failing that, an in-house replication on a smaller knowledge-worker sample is tractable.
- **Q4 (outside-frontier slope)** — Dell'Acqua BCG study (E3); HBS data release may include individual-level outcome grades plus inside/outside frontier classification per task. The linearity test of `u·(c_H − c_AI)` is a clean within-subjects design.
- **Q5 (workflow > capability)** — Everett 2025 (E8) demonstrates the workflow-restoration mechanism but does not directly test the bound. The cleanest test is a new RCT comparing routing strategies on the same `c_AI` (e.g., GPT-4 + naive flat-cyborg vs. GPT-3.5 + optimal routing on a matched task mix); existing data is suggestive, not decisive.
- **Q6 (explore-exploit on `c_AI`)** — Likely requires new experimental work or simulation. Closest analogs are the contextual-bandit and multi-armed-bandit-with-costly-verification literatures, but the knowledge-workflow application is novel; this is a Stage-4 originated study, not a re-analysis of existing data.

Stage 4's first move is to scope data availability for Q1-Q5; Q6 may require originating new data or a simulation harness. Following the human-psych-variation pattern, the Stage-4 build should live at `stage_outputs/technology-utilization-architecture/data/` with a curated CSV per fitting target, a runnable Python pipeline that reproduces every chart on the published `data.mdx`, and a `data/out/` folder for derived outputs.

## 12. Connections to other topics

Where this model attaches to sibling AI's-Research topics.

- **Human-psych-variation.** `λ` (skill-formation value) and `β` (atrophy rate) are individual differences. Need-for-Cognition (Buçinca 2021, E6) moderates `v` — high-NfC users verify more. Cognitive-style covariation belongs in that topic, not here.
- **Navigating-AI-world.** This model's `β` is the per-task version of nav-AI's `ΔM_comp` (competence erosion). The portfolio-level `S` aggregation is the within-work-domain version of nav-AI's `ΔV/ΔM` trade-off. The two models share the substitution-myth (L1) and verification-economics (G3) invariants; they differ in what they're optimising — nav-AI optimises a life-scale meaning budget, this model optimises a workday quality-per-attention budget.
- **Trust architecture (planned).** Sycophancy (E13) and human-side calibration on AI capability (O7) are trust-regime questions — what feedback signals make `c_AI` knowable.
- **Prediction & calibration (planned).** The topology's O7 connection. Calibration on `c_AI` is the calibration sub-problem this model treats as exogenous.
- **Information fidelity (planned).** `φ` (verification cost) depends on output-format quality and grounding — verifying a structured output with citations is cheap; verifying free prose is expensive. The information-fidelity topic should formalise what makes `φ` low or high.
- **Bedrock generating functions (planned).** The four-channel decomposition `V = Q − α·A + λ·S − σ·R` is a candidate generating-function pattern: every decision under attention scarcity has the same four channels. The bedrock topic should test whether this generalises beyond AI workflow.

## Glossary

- **autonomy level (`u`)** — fraction of a task's generation delegated to AI. Karpathy slider P2 made parametric.
- **verification depth (`v`)** — fraction of AI output independently checked by the human.
- **`c_H`, `c_AI`** — human and AI capabilities on a task type; probability of correct/high-quality output.
- **`c_⋆`** — verified-output ceiling; `max(c_H, c_AI)` under the assumption the human can verify.
- **`φ`** — verification-cost ratio: time-to-verify divided by time-to-generate-from-scratch.
- **`σ`** — stakes; weight on uncaught-error penalty.
- **`λ`** — skill-formation value of the task to the worker.
- **`α`** — attention price (utility weight on time). Normalised to 1.
- **`ε`** — residual attention at full delegation. The L1 substitution-myth invariant.
- **`β`** — per-task skill-atrophy rate under unverified delegation.
- **`M`** — per-task metacognitive routing tax. The G1 metacognitive-bottleneck invariant.
- **centaur** — Mollick: clean human/AI handoff with verification gate. `(u, v)` mid + high.
- **cyborg** — Mollick: interleaved sub-task delegation with partial verification. `(u, v)` mid + mid/low.
- **self-automator** — Randazzo: full delegation, no verification. `(u, v)` high + low. The atrophy-trap regime.
- **spec-driven / independent-then-synthesize** — Everett 2025; Compound Engineering: full delegation with full verification. `(u, v)` high + high.
- **do-yourself** — no AI involvement. `u ≈ 0`.
- **L1 (substitution myth)** — every offload creates new monitoring/verification work. Encoded as `ε > 0`.
- **L2 (joint surface)** — optimal allocation requires modelling the joint performance, not comparing solo capabilities. Encoded as the portfolio-level `argmax`.
- **L3 (parameterise by capability)** — the formalism is a generating function over θ, not a lookup table. The whole structure of the model.
- **G3 (verification trade-off)** — engagement is rational only when verification is cheap relative to expected payoff. The `−α·v·φ` term.
- **G7 (skill atrophy)** — capacities not exercised decay. The `−β·u·(1-v)` term.
- **G9 (generator-verifier asymmetry)** — production cost falls toward zero with AI; verification cost stays roughly constant. The asymmetry between `u·ε` (generation residual) and `v·φ` (verification full).

---

## Technology Utilization Architecture
*topic: technology-utilization-architecture · stage: overview · pass 0 · in-progress*

Optimal workflow architecture for an individual knowledge worker given the current AI / agent / automation toolset. Not "use AI" but the specific choreography — which tools for which cognitive operations, where human judgment is essential vs. bottleneck, and how to structure the feedback loops.

The topic running through the LLM Iterate pipeline. The question is not "does AI help" — that has been answered (15–55% productivity gains on well-bounded tasks, replicated across ~25 RCTs). The question is which workflow architecture maximizes output quality per unit of human attention, given that the binding constraint has shifted from production throughput to metacognitive load.

Stage 1 (lit review) maps three layers: a mature HCI / decision-science literature on appropriate reliance and complementarity; a classical foundation in cognitive systems engineering being re-imported (Bainbridge 1983, Klein et al. 2004, Hollnagel & Woods 2005); and a practitioner stack (Mollick, Karpathy, Anthropic, Cognition, Claude Code) that runs 12–18 months ahead of peer review. The headline finding across the empirical record is that workflow architecture predicts outcomes more reliably than which frontier model you use.

Stage 2 (topology) is the dependency graph — three foundational assumptions (attention-as-binding-constraint, jagged frontier exists, verification cost is comparable to generation cost) carry most of the inferential weight. Six crux nodes are where collapse propagates farthest. The graph also encodes how practitioner frameworks and academic theory map onto the same underlying structure.

Stages 3–5 will formalize the workflow choreography into a parameterized model (capability-by-operation × verification-cost × autonomy level → routing decision), test it against the available task-level evidence, and ship a small interactive tool for individual workflow design.

---

## Topology
*topic: technology-utilization-architecture · stage: topology · pass 9 · in-progress*

Dependency graph of the lit review. 66 nodes typed across nine classes (foundational assumptions / methods / empirical claims / logical necessities / generating mechanisms / synthesis / practitioner frameworks / open questions / distortions); seven cruxes (A1, A2, A3, A6, L1, L2, L3 — every weight-5 A node + every weight-5 L node) and four variant views (Vulnerability / Flow / Minimal / Capability-regime). The genuinely novel structural move for this topic: encoding the practitioner ↔ academic operationalization bridge that runs 12–18 months out of phase.

## TLDR

The lit review documents the research landscape on optimal human-AI workflow design. This topology asks the sharper question: **what depends on what?** Strip the field down to its load-bearing structure and the picture is surprisingly clean. Four foundational assumptions sit upstream of most of the empirical and synthesis nodes — that **human attention is the binding scarce resource** (A1), that **AI capability is heterogeneous across cognitive operations** (the jagged frontier; A2), that **verification cost is comparable to or cheaper than generation cost** (A3), and that **individual-level workflow optimization aggregates upward rather than being absorbed by organizational dynamics** (A6). If any one of them flipped, large regions of the picture would have to be rebuilt. Three logical guardrails (L1 substitution myth is wrong; L2 optimal allocation needs the joint performance surface, not solo accuracies; L3 parameterize by capability) cannot be falsified at all — they can only be ignored, which is exactly how most overconfident AI-rollout discourse proceeds. Everything else is methodology, empirical claim, generating mechanism, synthesis, practitioner framework, open question, or distortion vector.

The genuinely novel structural move for this topic is the **practitioner ↔ academic bridge** that the new node type **P** and the new edge type **op** make explicit. The practitioner stack (Mollick, Karpathy, Anthropic, Cognition, Claude Code) and the academic literature address the same structural problems but with different methodologies and on different timescales — and the relationship is bidirectional, not one-directional. Three patterns recur on the cross-community bridge. **(a) Practitioners ahead of academic measurement**: Mollick proposed the centaur/cyborg typology in 2024; Randazzo 2026 (HBS) is the first peer-reviewed empirical measurement of the behavior distribution (60/30/10) the typology names. **(b) Practitioners concretizing prior academic principles**: Karpathy's autonomy slider concretizes Shneiderman's 2D framework as a per-action user control surface; the AI Sandwich and Compound Engineering loop (Shipper / Every) concretize Tankelevitch's metacognitive-demands framework as a daily workflow practice. **(c) Practitioner and academic work converging in parallel without direct lineage**: Karpathy's autonomy slider (per-action UX) and Feng 2025's five-level academic taxonomy (per-task user roles) converged on the same discretized-levels-of-autonomy shape independently, at different granularities; Anthropic's agent design patterns (chain / route / parallelize / orchestrator-workers / evaluator-optimizer) address the same structural problem (joint allocation under capability heterogeneity) that L2D theory formalizes, but emerged from engineering practice rather than as L2D operationalizations. Note that some P-edges in the graph are *practitioner-internal* rather than cross-community — e.g., P6 (spec-driven development) → S4 (context engineering as central craft) and P7 (CLAUDE.md / context files) → S4 are both practitioner concrete-technique → practitioner-coined synthesis edges, not bridge edges, and the (a)/(b)/(c) classification doesn't apply to them. Reading the topology through this mixed bridge — rather than as a single integrated literature — disciplines the Stage-3 formalization: the model should encode invariants tracked across both communities (autonomy levels, verification gates, context structure, joint-allocation logic) parameterized by inputs the academic literature measures (capability gap, verification cost, skill-formation goals), without privileging either community as the source of authority.

The field's **weakest links** are not where the popular discourse focuses. Mainstream debate contests "does AI help" — but at the well-bounded-task level the productivity record is robust (E1: 15–55% gains across ~25 RCTs). The actual fragile zones in 2026 are: (a) **the aggregate-zero puzzle** (E4 / O2 attacks A6) — Humlum & Vestergaard's precise zero across 25,000 Danish workers tensions every micro-RCT result and is direct evidence against the individual-aggregation assumption that the entire individual-level frame depends on; (b) **whether the centaur regime persists** (E18 + S5) — Schoenegger's finding that even deliberately overconfident GPT improves forecasting suggests much of the gain comes from forced structured reasoning, not advice quality, raising the question of whether the centaur advantage survives a capability discontinuity; (c) **whether the frontier is mappable faster than it migrates** (O4) — a calibration race-condition the field hasn't tested; (d) **whether the binding reliance constraint is verification cost or verification calibration** (O7) — the design implication is different; (e) **long-term cognitive effects of continuous AI use** (O3) — Bastani's 17% unassisted-performance drop and the broader skill-atrophy literature point to a real risk but the longitudinal data window is still under twelve months for most studies.

This topology is the input to model formalization (Stage 3). The cleanest target is a **parameterized routing function**: for each (task type, capability profile, verification cost, autonomy level) tuple, produce an allocation decision that maximizes expected output quality per unit of human attention. The four variant views below (Vulnerability / Flow / Minimal / Capability-regime) read the same graph through different lenses to discipline that formalization choice — the capability-regime variant in particular sorts every node into stale-on-jump / structurally invariant / regime-dependent, which directly tells the model formalization which terms must be parameters (the regime-dependent ones) and which can be invariants (the stable ones).

## The graph

<CognitivePartnershipGraph client:load />

*Click a node for its claim and load-bearing weight; hover an edge for the relation type; drag to rearrange. The variant toggles read the same graph through different lenses.*

---

## How to read this graph

Every node in the lit review collapses to one of nine types. Edges between them carry one of seven relations. Together they make the structure inspectable.

### Node types

| Code | Type | What it is |
|---|---|---|
| **A** | Foundational assumption | A claim the field cannot operate without; if false, large downstream regions collapse |
| **M** | Methodological prerequisite | A study design or measurement approach that must work for the empirical claims to be testable |
| **E** | Empirical claim | A specific measured finding with an effect size and replication status |
| **L** | Logical necessity | Follows from definitions or algebra; not empirically refutable |
| **G** | Generating mechanism | A causal process that explains a pattern (metacognitive load, verification trade-off, ironies of automation) |
| **S** | Synthesis claim | An integrative statement combining multiple lower-level claims |
| **P** | Practitioner framework | A typology, slider, or pattern published by the practitioner stack ahead of academic formalization |
| **O** | Open question | Genuinely undecided with current methods or evidence |
| **D** | Distortion vector | Where motivated reasoning concentrates (typed by direction) |

### Edge types

| Code | Edge | Meaning |
|---|---|---|
| **dep** | depends-on | If target collapses, source collapses |
| **imp** | implies | Logical implication |
| **sup** | empirically-supports | Evidence relation |
| **conf** | confounds / inflates | Artifact relationship |
| **mod** | moderates | Changes magnitude |
| **op** | operationalizes | Practitioner framework concretizes an academic claim or vice versa |
| **corr** | corrects | Workflow architecture corrects naive allocation |
| **attacks** | attacks | Distortion vector targets a specific node |

### Weight scale (load-bearing weight, 1–5)

- **5** — crux node; collapse propagates across multiple sections of the lit review
- **4** — load-bearing within a section
- **3** — important but local
- **2** — corroborating
- **1** — decorative

---

## 1. Node catalog

Each node carries: type code · weight · short claim · key citation · status. Status flags: ✓ (robust/replicated), ~ (partial/qualified), ? (contested/open), ✗ (refuted, kept as historical reference).

### A — Foundational assumptions

| ID | Wt | Claim | Status |
|---|---|---|---|
| A1 | 5 | Human attention is the binding scarce resource — once production is automated, the bottleneck shifts upward to planning, evaluation, calibration. (Tankelevitch 2024) | ✓ |
| A2 | 5 | AI capability is heterogeneous across cognitive operations (the jagged frontier). (Dell'Acqua 2023/2025) | ✓ |
| A3 | 5 | Verification cost is comparable to or cheaper than generation cost — otherwise rational engagement collapses. (Vasconcelos 2023; Fok & Weld 2023) | ~ |
| A4 | 4 | Knowledge work decomposes into sub-tasks that can be selectively delegated. (Vaccaro 2024 sub-task finding) | ✓ |
| A5 | 4 | The frontier has stable topological features even as the boundary shifts — i.e., it is mappable. | ? |
| A6 | 5 | Individual-level workflow optimization aggregates upward — gains aren't fully absorbed by organizational dynamics (review bottlenecks, managerial reabsorption, coordination costs). The load-bearing assumption behind the entire individual-level framing. Humlum-Vestergaard aggregate-zero is direct evidence against. | ? |

### M — Methodological prerequisites

| ID | Wt | Claim | Status |
|---|---|---|---|
| M1 | 5 | Randomized controlled trials of AI-augmented work. (~25 published 2023–2026) | ✓ |
| M2 | 4 | Strategy-graded reliance metrics (vs. outcome-graded). (Vasconcelos / Fok & Weld pivot) | ✓ |
| M3 | 4 | Field-deployed measurement (real repos, real meetings). (METR 2025) | ~ |
| M4 | 3 | Telemetry / log-based behavioral observation. (Mozannar CUPS 2024) | ✓ |
| M5 | 3 | Cognitive Task Analysis (CTA, ACTA). (Crandall, Klein & Hoffman 2006) | ~ |

### E — Empirical claims

| ID | Wt | Claim | Status |
|---|---|---|---|
| E1 | 5 | 15–55% productivity gains on well-bounded knowledge tasks. (Brynjolfsson 2023; Noy 2023; Peng 2023; Cui 2025) | ✓ |
| E2 | 5 | Skill-leveling: novices gain most, top performers near-zero — on well-defined tasks. (Brynjolfsson +34%/0%) | ✓ |
| E3 | 5 | Outside-frontier AI use causes a 19-pp quality drop. (Dell'Acqua BCG study) | ✓ |
| E4 | 5 | Aggregate labor-market effects are zero at 2-year horizons. (Humlum & Vestergaard 2025, 25k workers) | ✓ |
| E5 | 4 | Explanations alone don't yield complementary performance — they increase acceptance regardless of correctness. (Bansal 2021) | ✓ |
| E6 | 4 | Cognitive forcing reduces overreliance — but only for high-Need-for-Cognition users. (Buçinca 2021) | ✓ |
| E7 | 4 | On naive workflows, AI alone outperforms human+AI in expert domains. (Goh 2024 JAMA NO) | ✓ |
| E8 | 5 | "Independent-then-synthesize" workflow restores complementarity in the same domain. (Everett 2025) | ✓ |
| E9 | 4 | Unguardrailed AI in learning produces 17% worse unassisted post-session performance. (Bastani PNAS 2025) | ✓ |
| E10 | 3 | Behavior distribution: ~60% cyborg, ~30% centaur, ~10% self-automator. (Randazzo HBS 26-036) | ✓ |
| E11 | 4 | LLM capability decays sharply up Bloom's hierarchy: ~81% Remember → 13–41% Analyze. (Ma 2025) | ✓ |
| E12 | 4 | Verification time is a substantial fraction of total interaction in coding contexts. (Mozannar CUPS 2024) | ✓ |
| E13 | 4 | Sycophancy escalation: AI flips correct human judgments to incorrect on pushback. (Randazzo HBS 26-021) | ✓ |
| E14 | 4 | Multi-agent orchestrator-worker outperforms single-agent +90.2% at ~15× token cost. (Anthropic 2024/2025) | ✓ |
| E15 | 4 | Routing > model selection — task-to-tool dispatch beats single-best-model. (Paterson 2026) | ✓ |
| E16 | 3 | Higher AI confidence correlates with less critical thinking enacted. (Lee et al. CHI 2025) | ~ |
| E17 | 4 | Read/write asymmetry: multi-agent works for read-heavy tasks; breaks on write-heavy unless writes are serialized. (Cognition / Anthropic / LangChain consensus) | ✓ |
| E18 | 3 | Both calibrated AND deliberately overconfident GPT assistants improve human forecasting +23–43%. (Schoenegger 2024/2025) | ~ |

### L — Logical necessities

| ID | Wt | Claim | Status |
|---|---|---|---|
| L1 | 5 | Substitution myth is wrong — every offload creates new monitoring/verification/coordination work. (Dekker & Woods 2002; Bainbridge 1983) | ✓ |
| L2 | 5 | Optimal allocation requires modeling the *joint* performance surface, not solo accuracies. (Madras / Mozannar L2D theory) | ✓ |
| L3 | 5 | Allocation model must be parameterized BY capability, not depend on FIXED capability — generating function vs. lookup table. | ✓ |
| L4 | 4 | High automation and high human control can coexist. (Shneiderman 2D) | ✓ |

### G — Generating mechanisms

| ID | Wt | Claim | Status |
|---|---|---|---|
| G1 | 5 | Metacognitive bottleneck — load shifts from production to planning, evaluation, calibration. (Tankelevitch 2024) | ✓ |
| G2 | 5 | Ironies of automation — AI handling routine work degrades human capacity to catch rare critical errors. (Bainbridge 1983; Simkute 2024) | ✓ |
| G3 | 5 | Verification-cost trade-off — engagement is rational only when cheap. (Vasconcelos 2023) | ✓ |
| G4 | 5 | Jagged frontier mechanism — capability heterogeneous; the boundary is personal and dynamic. | ✓ |
| G5 | 4 | Correlation neglect — humans treat AI advice as independent evidence despite shared training data. (Amin 2026) | ~ |
| G6 | 3 | Cognitive forcing — committing to a view first breaks anchoring. | ✓ |
| G7 | 4 | Skill atrophy — capacities not exercised decay. | ✓ |
| G8 | 4 | Cross-task productivity bundling — speedups bottlenecked by linked tasks. (Cowen 2026) | ~ |
| G9 | 4 | Generator-verifier asymmetry — production cheap, checking expensive. (Karpathy) | ✓ |
| G10 | 3 | Multi-tool attention interference — Wickens' Multiple Resource Theory predicts tool-stack overload (Cursor + ChatGPT + meeting incurs cost beyond the sum of per-tool costs). Mechanism well-established; absent from AI workflow research. | ~ |

### S — Synthesis claims

| ID | Wt | Claim | Status |
|---|---|---|---|
| S1 | 5 | **Workflow architecture predicts outcomes more reliably than model capability.** | ✓ |
| S2 | 5 | The optimization target is metacognitive efficiency, not throughput. | ✓ |
| S3 | 4 | Knowledge work shifting from production to critical integration. (CHI 2025) | ✓ |
| S4 | 4 | Context engineering is the central craft. (Anthropic / Cognition consensus) | ✓ |
| S5 | 3 | Centaur advantage may be a transient regime. | ? |

### P — Practitioner frameworks

| ID | Wt | Claim | Status |
|---|---|---|---|
| P1 | 5 | Mollick centaur / cyborg / self-automator typology. | ✓ |
| P2 | 4 | Karpathy autonomy slider. | ✓ |
| P3 | 4 | Anthropic agent design patterns (chain / route / parallelize / orchestrator-workers / evaluator-optimizer). | ✓ |
| P4 | 4 | Cognition read/write asymmetry as agent-system principle. | ✓ |
| P5 | 3 | Compound Engineering / AI Sandwich (Shipper, Every). | ~ |
| P6 | 3 | Spec-driven development (spec → plan → execute). | ~ |
| P7 | 3 | Personal context files (CLAUDE.md / system-prompt patterns). | ✓ |

### O — Open questions

| ID | Wt | Claim | Status |
|---|---|---|---|
| O1 | 5 | Does AI help experts on open-ended judgment? | ? |
| O2 | 5 | Why are aggregate effects zero given the micro-RCT record? | ? |
| O3 | 5 | Long-term cognitive effects under continuous AI use. | ? |
| O4 | 4 | Frontier migration vs. calibration rate. | ? |
| O5 | 4 | Right unit of analysis — individual vs. team vs. value chain. | ? |
| O6 | 3 | Centaur taxonomy durable, or interface artifact? | ? |
| O7 | 4 | Verification cost vs. verification calibration as binding constraint. | ? |

### D — Distortion vectors

| ID | Wt | Claim | Targets |
|---|---|---|---|
| D1 | 4 | AI-maximalist distortion — RCT gains read as aggregate revolution; ignores Humlum-Vestergaard zero. | E4, S1, O2 |
| D2 | 4 | Productivity-only distortion — counts speed gains, ignores skill atrophy and metacognitive load. | G7, S2, E9 |
| D3 | 3 | "Just use the best model" distortion — ignores routing finding (E15) and architecture-over-capability evidence. | E15, S1, S4 |
| D4 | 3 | Practitioner-only distortion — dismisses formalization as category error. | L3, S1 |

---

## 2. Edge catalog (key chains, not exhaustive)

**Foundation → Method.** A2 → M1 (RCTs reveal the jagged frontier); A3 → M2 (strategy-graded metrics measure verification cost); A1 → M3 (field deployment shows real attention budget).

**Method → Empirical.** M1 produces the productivity record (E1–E9). M2 → E5 (Bansal explanations don't help once strategy-graded). M2 → E12 (CUPS is the strategy-graded measure of coding). M3 → E2 (METR shows experts gain less in real repos).

**Mechanism → Empirical.** G1 → E12 (metacognitive load shows up as verification time). G2 → E13 (sycophancy is the rare critical error ironies-of-automation predicts will get missed). G3 → E5 (rational verification trade-off explains why explanations fail). G4 → E3 (jagged frontier produces outside-frontier harm). G7 → E9 (skill atrophy → unassisted-performance drop). G8 → E4 (cross-task bundling explains aggregate zero). G10 → E10 (cyborgs interleaving multiple tools should incur higher MRT interference than centaurs handing off discrete sub-tasks — testable but unmeasured prediction).

**Empirical → Synthesis.** E1, E2, E3, E8 → S1 (workflow architecture > model capability — the integrated headline). E12, E16 → S2 (metacognitive efficiency target). E11 → S3. E14, E15, E17 → S4 (context engineering as central craft). E18 → S5.

**Empirical → Open.** E2, E4 → O2 (the aggregate puzzle). E2 → O1 (helps experts?). E9 → O3 (long-term cognitive). E1, E3 → O4 (frontier migration). E4 → O5 (right unit). E13 → O7 (calibration vs. cost).

**Logical guards.** L1 → G2 (substitution myth → ironies of automation). L2 → S1 (joint surface needed). L3 → S4 (parameterize-by-capability is what makes context engineering generalize). L4 → P2 (Shneiderman 2D legitimates the autonomy slider).

**Practitioner ↔ Academic (the central conceptual move of this topology, encoded as op edges).** The relationship type varies — see TLDR para 2 for the (a)/(b)/(c) classification of cross-community bridge edges. P1 → E10 (Mollick named the typology; Randazzo 2026 measured the behavior distribution it predicts — pattern (a), academia retrospectively measures). P2 → L4 (Karpathy's autonomy slider concretizes Shneiderman's 2D framework as a per-action user control surface — pattern (b), prior academic principle made tangible), with the reverse edge L4 → P2 (Shneiderman legitimates the slider design). P3 → L2 (Anthropic's agent design patterns and Madras / Mozannar L2D theory address the same structural problem — joint allocation under capability heterogeneity — but emerged independently — pattern (c), parallel convergence without direct lineage). P4 → E17 (Cognition's read/write principle IS the design-actionable form of the read/write asymmetry finding — pattern (a) inverted: practitioners stated and measured it together). P5 → S2 (compound engineering / AI Sandwich applies Tankelevitch's metacognitive-demands framework at the workflow level — pattern (b)). **P6 → S4 and P7 → S4 are practitioner-internal, not bridge edges**: S4 (context engineering as central craft) is itself a practitioner-coined synthesis (Anthropic / Cognition consensus), and P6 (spec-driven development) and P7 (CLAUDE.md / personal context files) are concrete practitioner techniques that operationalize that practitioner synthesis. The (a)/(b)/(c) classification doesn't apply because both endpoints sit in the practitioner community.

**Foundation → Foundation (the project-frame edges).** A6 → S1 (S1 is meaningful only if individual optimization aggregates). E4 → A6 (aggregate-zero is the direct attack on A6). O2 → A6 and O5 → A6 (the open questions whose resolution will close A6 either way).

**Distortion attacks.** D1 → S1, E4 (treats RCT gains as aggregate proof; ignores Humlum). D2 → G7, S2, E9 (counts speed; ignores atrophy). D3 → E15, S1, S4 (ignores routing). D4 → L3, S1 (denies formalization possibility).

---

## 3. High-stakes nodes — by structural role

Six categories of structural role, sorted by how their failure modes propagate. The single most useful conceptual move is keeping the cruxes (inputs) separate from the headline (an output) and from the reframers (mechanisms whose magnitude is open). Conflating these three under a single label of "important findings" produces most of the bad-faith debate around AI workflow design.

### 3a. Foundational cruxes — collapse rebuilds regions

These are the **input assumptions** the entire individual-level framing rests on. Falsification doesn't change interpretation; it forces rebuilding. All four are weight-5 foundational-assumption nodes (the A class).

- **A1** (human attention is the binding scarce resource). If false, throughput-optimization wins and the entire metacognitive-bottleneck framing dissolves; design priorities flip back toward maximum AI delegation. The metacognitive-bottleneck mechanism G1 is the *consequence* of A1 + production-automation, not a separate axiom — which is why G1 isn't itself a crux.
- **A2** (jagged frontier — AI capability is heterogeneous across operations). If a capability discontinuity produced uniformly-good AI across all knowledge-work operations, the "map your frontier" design imperative collapses and outside-frontier harm (E3) dissolves.
- **A3** (verification cost is comparable to or cheaper than generation cost). If verification became prohibitively expensive — AI outputs so complex or fast-moving that human checking is intractable — the rational-cost-benefit reframing of overreliance (G3) dissolves and the design problem becomes "trust without verification."
- **A6** (individual-level optimization aggregates upward). If gains are absorbed by organizational dynamics — review bottlenecks downstream, managerial reabsorption, coordination costs — individual workflow optimization is locally optimal but globally insufficient. Humlum-Vestergaard's aggregate zero is direct evidence against. **This is the project-frame crux**: the entire individual-level optimization target only matters if A6 holds, and its status is genuinely open.

### 3b. Logical guardrails — unfalsifiable, ignored at peril

Cannot be falsified — only ignored. All three are weight-5 logical-necessity nodes (the L class).

- **L1** (substitution myth is wrong — every offload creates new monitoring/verification/coordination work). Bainbridge 1983 / Dekker & Woods 2002. The structural property of any principal-agent delegation; AI-rollout discourse routinely treats it as if it could be ignored.
- **L2** (optimal allocation requires modeling the *joint* performance surface, not solo accuracies). The Madras / Mozannar L2D formal result. A mathematical property of how joint performance combines — cannot be falsified empirically. The most common practitioner shortcut — "use AI when AI is better, use human when human is better" — implicitly assumes solo accuracies suffice; ignoring the variance and correlation structure of the two agents' errors is exactly the move L2 forbids. The agent-design patterns (P3) that *do* work are the ones that respect L2 by construction.
- **L3** (allocation model must be parameterized by capability, not depend on fixed capability). The generating-function-vs-lookup-table commitment. Practitioner-only frameworks routinely violate this by hardcoding "GPT-4 is good at X, bad at Y" — the pattern goes stale within months.

### 3c. Reframer mechanisms — magnitude is the live question

High-weight non-crux mechanism nodes whose *magnitude* (not existence) is what reshapes interpretation. Each is well-supported as a phenomenon; the open question is the share of variance they explain.

- **G1** (metacognitive bottleneck). CHI 2024 Best Paper finding; robust phenomenologically. The open magnitude question: what share of the labor force is in the regime where G1 is binding? For workers still production-constrained, G1 is premature; for power users, it is binding. The size of each population determines whether the metacognitive-efficiency target (S2) is the right design priority for "individual workers" generically or only for a subset.
- **G3** (verification-cost trade-off). Vasconcelos / Fok-Weld reframing. The mechanism is real; the open question is whether it captures most of the variance in reliance behavior, or whether verification *calibration* (O7) is doing the rest of the work.
- **G4** (jagged frontier mechanism). The capability-heterogeneity story is robust; the open magnitude question is the rate at which the boundary shifts (O4) and whether the topological features are stable (A5).

### 3d. Headline conclusion — a synthesis output, not a crux

- **S1** (workflow architecture > model capability). The integrated finding the topology is organized around. **S1 is what the cruxes plus mechanisms produce, not a load-bearing input.** It is a weight-5 synthesis node, but distinguishing "headline conclusion" from "crux" matters: cruxes are inputs whose falsification rebuilds the graph; conclusions are outputs whose falsification only means the rebuild was downward (the conclusion was wrong) rather than upward (an input was wrong).

### 3e. Corroborating / illustrative

Two senses of "droppable" need to be distinguished here. Some nodes can be removed without breaking S1's *conclusion* but are load-bearing as *exemplars* of the topology's classification structure (especially the practitioner ↔ academic bridge in TLDR para 2); others are droppable in both senses.

- **Load-bearing as exemplars, droppable for S1**: E10 (60/30/10 distribution — the canonical (a) bridge example, paired with P1 to demonstrate "practitioners ahead of academic measurement"; if removed, S1 still holds but the (a) example loses its empirical anchor), P5 (compound engineering / AI Sandwich — the canonical (b) bridge example, paired with S2 to demonstrate "practitioners concretizing prior academic principles"; if removed, S1 still holds but the (b) example weakens).
- **Droppable in both senses**: E18 (Schoenegger overconfident-AI-still-helps — interesting but tangential to S1 and not used as a bridge example), G6 (cognitive forcing as mechanism — local, not load-bearing for S1 or the bridge framing), P6 (spec-driven development — practitioner-internal P → S4 edge, not a bridge example, not load-bearing for S1).

### 3f. Distortion vectors

D1–D4 are pedagogically intentional, not decorative. Each names a real motivated-reasoning pattern and the specific empirical/logical claims it targets. Distortions are useful for readers calibrating where their own priors might be selecting against the evidence.

---

## 4. Weakest links

Where the graph is genuinely fragile in 2026, ranked by potential propagation if the link breaks:

1. **A6 / O2 / O5 (the aggregate-zero / unit-of-analysis cluster).** The most consequential weakness in the graph. Humlum-Vestergaard's precise zero across 25,000 Danish workers is direct evidence that individual workflow gains may not aggregate; if true, A6 falsifies, the project-frame inverts, and individual workflow optimization becomes locally optimal but globally insufficient. Possible mechanisms (cross-task bundling per Cowen 2026, organizational reabsorption, weak wage pass-through) are testable but untested. Until O2 is resolved, every claim about *organizational* benefit downstream of S1 is contingent.
2. **O3 (long-term cognitive effects).** Bastani's 17% drop is a single-session study with a short post-session window — the lit review documents the in-session vs. afterward contrast but no specific durability timescale. If a 2-year longitudinal RCT confirms broad skill atrophy across cognitive operations, the design implication becomes "AI-free zones" at much larger scale than current practice — and the productivity-only distortion (D2) goes from intellectually wrong to materially harmful.
3. **O4 (frontier migration vs. calibration rate).** If model capabilities shift faster than humans can recalibrate their personal frontier maps, the "map your jagged frontier" design imperative becomes unfollowable. The likely partial resolution is that the frontier has stable topological features (A5) even when the boundary moves, but A5 is empirically untested.
4. **O7 (verification cost vs. calibration as binding constraint).** Reducing verification cost helps a rational actor; improving calibration helps a miscalibrated one. The interventions are different. The sycophancy finding (E13) suggests calibration is a real second binding constraint.
5. **S5 / O6 (centaur regime persistence).** If a capability discontinuity makes AI better than humans on the full task, the centaur typology becomes historical.
6. **A5 (frontier mappable).** Empirically untested. If false, individual-level workflow design is structurally impossible at the per-task granularity the practitioner stack assumes.

---

## 5. Variants

Each variant reads the same graph through a different lens.

### Variant A — Vulnerability (where does this break?)

Highlights the seven cruxes (A1, A2, A3, A6 foundational; L1, L2, L3 logical guardrails) plus the weight-5 nodes downstream. If any crux flips, propagation is concentrated through this subgraph. Useful for stress-testing: pick a crux, imagine it inverts, and trace the consequences through the highlighted subgraph.

### Variant B — Flow (how does causation propagate?)

Restricts to the A → M → E → S/G cascade plus practitioner operationalizations (P → E/L/S via op edges). The "what generated what" view: foundational assumptions enable methods, methods produce empirical findings, mechanisms explain them, syntheses integrate, and practitioner frameworks operationalize the resulting design implications.

### Variant C — Minimal claim set

Smallest set of claims that still yields the headline conclusion (S1: workflow architecture > model capability). Approximately: A2 + A3 + E3 + E8 + L2 + G3 + G1 + S1. Eight nodes. Removing any one breaks the qualitative shape.

### Variant D — Capability-regime fragility (the topic-specific variant)

Which nodes go stale if frontier capability jumps? The central worry the lit review's adversarial section names. Of the 66 nodes, 35 sort into one of three regime-fragility classes; the rest (methods, supporting empirical findings, distortions) are regime-orthogonal. The classification:

- **Stale-on-jump (6 nodes — likely to invert).** E2 (skill-leveling: if AI exceeds top performers, the +34/0 pattern flips). E3 (outside-frontier harm: dissolves if frontier becomes uniform). E10 (60/30/10 behavior distribution: regime-bound to current tool affordances). E18 (overconfident-AI-still-helps: if AI is reliably correct, the "structured reasoning is the gain" reading dissolves). S5 (centaur transience: literally about transitioning out). P1 (Mollick taxonomy: becomes historical the way the chess centaur literature reads as historical now).

- **Stable-on-jump (18 nodes — structurally invariant).** All four logical necessities: L1 (substitution myth), L2 (joint performance surface), L3 (parameterize-by-capability), L4 (autonomy + control coexist). The mechanism invariants: G1 (metacognitive load just shifts to new decisions), G2 (ironies of automation are a property of any principal-agent delegation), G3 (verification trade-off — the threshold moves but the shape is invariant), G9 (generator-verifier asymmetry), G10 (multi-tool attention interference — Wickens MRT is a property of human cognitive architecture, not of AI capability). The foundational A1 (attention-as-binding-constraint is a fact about humans, not about AI). The synthesis claims that follow from the invariants: S2 (metacognitive efficiency target), S3 (production → critical integration), S4 (context engineering as central craft). The control-surface practitioner patterns: P2 (autonomy slider), P3 (Anthropic agent design patterns), P5 (compound engineering), P6 (spec-driven development), P7 (CLAUDE.md context files).

- **Regime-dependent (11 nodes — depends on which way capability jumps).** Foundational A2 (jagged frontier could become smooth or fragment into a different jagged shape), A3 (verification cost could fall further or rise as outputs become more complex), A5 (frontier-mappability depends on rate of migration), A6 (aggregation could go either way as tools get integrated). The headline S1 (workflow > capability could invert if a capability gap dominates). The high-leverage empirical claims: E14 (multi-agent could become moot if single-agent matches), E15 (routing depends on heterogeneity A2), E17 (read/write asymmetry depends on whether AI handles serial writes natively). The mechanism G4 (jagged frontier follows A2). The open question O4 (frontier migration rate IS the regime question).

The model formalization should be stable under capability change — meaning it should be parameterized by the L1/L2/L3/L4 + G1/G2/G3/G9 + A1 invariants and treat A2/A3/A6/E14/E15/E17 as inputs that vary by capability regime. The 6 stale nodes are the ones the formalization should *not* hardcode.

---

## 6. Stage-3 handoff

This topology is the input to model formalization. The cleanest target is a **parameterized routing function**:

```
delegate(task, worker, AI, context) → action
```

where the action is one of `{do_yourself, consult_AI, delegate_to_AI, refuse}`, and the function is parameterized by:

- **Task type** (per Bloom hierarchy + cognitive task analysis decomposition)
- **Worker capability profile** on that task type (via prior frontier mapping)
- **AI capability profile** on that task type (via per-task benchmark or recent personal experience — the L3 invariant means this is an input, not a hardcoded constant)
- **Verification cost** (function of task type and output format)
- **Stakes / reversibility** (high stakes → independent-then-synthesize per E8)
- **Skill-formation goal** (if the task is one whose capability the worker wants to maintain → guardrail mode per E9)

The `stage_outputs/<topic>/<stage>.md` folder convention holds for this topic too: raw working drafts live in `stage_outputs/technology-utilization-architecture/<stage>.md`; polished versions move into `src/content/ai_research/technology-utilization-architecture/<stage>.mdx`. So far the folder contains the lit review and this topology draft; subsequent stages will accumulate there.

This topology also feeds three natural sibling topics down the line. **Navigating an AI World** is the structural / civilizational view that holds the individual-level optimization in tension with the organizational-level disruption (its own Crux 5 / A6 sits next to mine). **AI Cognitive Profile** is the orthogonal view: rather than asking "how should an individual route tasks," ask "where does AI capability diverge from human capability across the O*NET task taxonomy" — that topic supplies what this topic treats as a black-box input (the per-task capability gradient) and inversely, this topic supplies what that topic treats as a black-box (what individuals should *do* given the gradient). **Prediction and Calibration** is the natural attachment point for O7 (verification cost vs. verification calibration as binding constraint): if calibration on AI error rates is the second binding constraint behind verification cost, the calibration topic is where that gets formalized. The Stage-3 model formalization should leave clean attachment points for all three.

---

## 7. Next moves — three Stage-3 options

Three formalization paths, each with pros / cons / Stage-4 implications.

### Option A — Capability × verification-cost dispatch table (decomposition)

**What it is.** Formalize the routing function as a 2D table: capability gap (worker minus AI on this task) × verification cost ratio (verify-cost / generate-cost). Four quadrants → four allocation rules.

- **Pros.** Maps cleanly onto the existing RCT record (quadrants align with Brynjolfsson, Dell'Acqua, METR, Everett). Easy to visualize. Practitioner-actionable.
- **Cons.** Risks violating L3 (looks like a lookup table). Mitigated if the table is generated from inputs rather than hardcoded.
- **Stage-4 implication.** Test against the published RCTs by classifying each into a quadrant.

### Option B — Generator-verifier loop with autonomy slider (generating function) [recommended]

**What it is.** Formalize the workflow as a recurrent loop: at each step, the worker chooses an autonomy level (Karpathy P2 / Feng 2025 five-level operator → observer). The loop has a verification gate; the gate's strictness is a parameter. Output: an interactive dashboard where the user inputs (task type, capability gap, verification cost, stakes, skill-formation goal) and gets a recommended autonomy level + verification cadence.

- **Pros.** Directly operationalizes Karpathy's autonomy slider (P2) using Shneiderman 2D (L4) and Vasconcelos verification economics (G3). Survives capability change (L3-compatible). Naturally maps onto an interactive site component.
- **Cons.** Requires choosing a parameterization of "verification cost" that holds across task types — non-trivial.
- **Stage-4 implication.** Test against Bastani's guardrail RCT (verification gate stringency moderating skill atrophy), Everett's independent-then-synthesize (a specific autonomy-cadence schedule), and the Mollick centaur/cyborg empirical distribution (which loops people actually run).

### Option C — Principal-agent with imperfect agent (mechanism design)

**What it is.** Apply contract-theory machinery (asymmetric information, monitoring vs. trust trade-off) to the single-worker / AI-tool relationship.

- **Pros.** Formally rigorous. Connects to a mature economic literature.
- **Cons.** Heavyweight; principal-agent assumes strategic agents, but LLMs are imperfect-but-non-strategic.
- **Stage-4 implication.** Hardest to validate against the existing RCT record.

**Recommendation: Option B.** The generator-verifier loop with an autonomy slider is the most directly testable, the most actionable as an interactive site artifact (Stage 5), and the most clearly L3-compatible. Specifically, Option B engages the regime-stable invariants identified in Variant D — L1 (every offload creates new monitoring work, baked into the loop's verification gate), L2 (joint-surface allocation, baked into the autonomy-level choice), L3 (parameterized by capability inputs rather than hardcoded), G3 (verification trade-off, the parameter that sets gate strictness), G9 (generator-verifier asymmetry, the loop's central asymmetry), and G10 (multi-tool attention interference — the loop should track concurrent-tool load as a Wickens-MRT input, not just per-task parameters; "running three coding agents in parallel" is a different regime from "running one") — while taking A6 (individual optimization aggregates) as the explicit assumption whose falsification would mean the loop is locally optimal but globally insufficient. Option A can be a sub-component of Option B (the dispatch table determines the autonomy-level choice). Option C is held in reserve.

---

## 8. Objections to this topology

**Objection 1: "The practitioner / academic split is overstated — the field is actually more integrated than the typing P-vs-S suggests."** Steelman: many researchers (Mollick, Karpathy, Tankelevitch via MSR's Tools for Thought) span both communities; some practitioner posts (Anthropic's effective agents) cite academic literature; the typing risks reifying a divide that's already partial. Response: the *people* span both, but the *artifacts* arrive on different timescales and through different processes. Mollick proposed the centaur/cyborg typology in 2024; Randazzo 2026 (HBS) is the first peer-reviewed empirical measurement of the distribution. Karpathy's autonomy slider was articulated as a tool-design concept (in his 2024 Software 1.0/2.0/3.0 talks and Cursor's Tab → Cmd+K → Agent Mode UX) about a year before Feng 2025's five-level academic taxonomy converged independently on a similar discretized-levels structure — neither grounds the other; they are parallel work on the same concern at different granularities (Karpathy's is per-action, Feng's is per-task user role). Anthropic's agent design patterns and L2D theory likewise address the same structural problem with no direct lineage. The typing isn't claiming a clean division; it's making the *structurally different roles* visible — empirical measurement, formal-theory derivation, engineering-practice synthesis — so we don't conflate "Anthropic shipped a pattern that works" with "L2D theory predicts this pattern is optimal."

**Objection 2: "The 7-crux selection is biased toward the framing the lit review wants."** Steelman: a critic could argue capability-first cruxes (e.g., "Bloom-level decay determines all task allocation") are equally defensible, or that the lit-review-driven framing inherits whatever bias the lit review has. Response: the crux set is **structurally typed**, not picked by impact. The criterion is "claims whose collapse rebuilds regions of the graph," and that maps cleanly onto exactly two node classes — foundational assumptions (A) and logical necessities (L). The cruxes are A1, A2, A3, A6 (every weight-5 A node) plus L1, L2, L3 (every weight-5 L node). No mechanism (G), synthesis (S), empirical (E), practitioner (P), or open (O) node is a crux, because their structural role is downstream of the cruxes — mechanisms explain, syntheses integrate, empirical findings test, practitioner frameworks operationalize, open questions sit at the frontier. The headline conclusion S1 (workflow > capability) is *not* a crux even though it is weight-5; it's an output of the graph, and excluding it from the crux set is the discipline that distinguishes "what the graph rests on" from "what the graph concludes." A capability-first alternative would have to add a foundational A node like "Bloom-level decay is the dominant capability gradient" — which the lit review doesn't support; LLMs decay sharply on Bloom, but task allocation is also shaped by verification cost, autonomy level, and skill-formation goals. The selection is therefore lit-review-driven, but the *typing rule* (A + L only, every weight-5) is independent of the lit review's framing — it is just "which nodes carry the structural role of being inputs the rest of the graph rests on."

**Objection 3: "The capability-regime fragility variant is a hedge — it concedes the project will go stale."** Steelman: Variant D essentially admits that half the empirical claims could invert under a capability jump; if so, why formalize at all? Response: Variant D is the discipline that *justifies* formalization. The point is to identify which nodes are stable-under-jump (the L1-L3-G1-G2-G3 invariants) and build the formalization on those. Practitioner heuristics, by contrast, are not capability-stable by design — they update faster but go stale faster too.

**Objection 4: "The aggregate-zero puzzle (E4 / O2) is so consequential that the entire individual-level framing might be wrong, and the topology should pivot to the organizational level instead."** Steelman: if Humlum-Vestergaard's zero is the true-population result, individual workflow optimization is rearranging deck chairs. Response: this is exactly what A6 (now a foundational crux) and O5 (right unit of analysis) name. The honest position is that the individual frame is *one valid level* of analysis — design-actionable for the worker who controls their own workflow — and the organizational frame is a *separate* level that should be a sibling artifact, not a replacement.

**Objection 5: "Topology + model formalization is over-engineering for a moving target. Practitioner heuristics adapt faster than academic formalization can ship; by the time you finish the model, the field has moved. The honest move is to skip directly to a build artifact using current best practices."** Steelman: this is real. The lit review already documents that the practitioner stack runs 12–18 months ahead of peer review precisely because it doesn't try to formalize. Cursor, Claude Code, and the Anthropic agent-design patterns are already shipping; users adapting heuristics in real time will outperform users waiting for an academic model. The L3 invariant ("parameterize by capability") is meant to address this, but it's a hope, not a guarantee — a formalization built on capability-stable invariants might *still* miss the regime where capability discontinuity dominates everything. Response: even granting all of that, the topology IS the artifact whose stable invariants survive capability change. Variant D identifies an 18-node regime-stable subgraph: all four logical necessities (L1 substitution myth, L2 joint surface, L3 parameterize-by-capability, L4 autonomy + control coexist), the five mechanism invariants (G1 metacognitive load shifting, G2 ironies of automation, G3 verification trade-off, G9 generator-verifier asymmetry, G10 multi-tool attention interference via Wickens MRT), the foundational A1 (attention-as-binding-constraint is a fact about humans, not AI), the synthesis claims that follow from the invariants (S2 metacognitive efficiency target, S3 production → critical integration, S4 context engineering as central craft), and the control-surface practitioner patterns (P2 autonomy slider, P3 Anthropic agent design patterns, P5 compound engineering, P6 spec-driven development, P7 CLAUDE.md context files) — none of which are capability-bound claims. Practitioner heuristics adapt faster but lossy: they don't preserve the reasoning behind the rule, so users can't tell when the heuristic stops applying. The 12–18 month lag goes both ways — practitioners ship faster, but they also rediscover Bainbridge 1983 forty years late. The model formalization's value is less "predict optimal allocation" than "encode the invariants explicitly so the next capability shift doesn't require re-litigating from scratch." That said: the objection has bite for Stage 5 specifically. If the build artifact is a fragile prediction tool that hardcodes 2026 capability, it goes stale. If it is a *frame* (the autonomy slider, the verification-cost trade-off visualized) that the user fills in with their own current capability profile, it is L3-compatible and survives.

---

## 9. Glossary

- **AI Sandwich** — Shipper / Every: humans frame the task and review the output; AI handles the middle.
- **autonomy slider** — Karpathy: a per-action user control surface (Cursor: Tab → Cmd+K → Agent Mode); operationalizes Shneiderman's 2D framework.
- **Bloom's taxonomy** — Remember → Understand → Apply → Analyze → Evaluate → Create. LLM capability decays sharply going up.
- **centaur** — Mollick: clean human/AI role separation; discrete handoffs based on per-task frontier mapping.
- **CoALA** — Cognitive Architectures for Language Agents (Sumers 2024). Maps LLM agents to ACT-R / SOAR memory + action structures.
- **complementarity potential vs. team performance** — Hemmer 2025 distinction: mathematical possibility vs. actual achievement of human+AI exceeding either alone.
- **compound engineering** — Shipper / Every: plan → work → review → compound. Each loop's outputs feed the next loop's inputs.
- **context engineering** — Anthropic / Cognition: the discipline of constructing what the model sees. Practitioner consensus: higher leverage than model selection.
- **CTA / ACTA** — Cognitive Task Analysis / Applied CTA. Method for decomposing what a knowledge worker actually does cognitively.
- **CUPS** — Cognitive Use Pattern States; Mozannar 2024 telemetry-grounded taxonomy of programmer-Copilot interaction.
- **cyborg** — Mollick: continuous interleaving of human and AI at sub-task granularity. The empirical majority pattern.
- **generator-verifier asymmetry** — production cost falls toward zero with AI; verification cost stays roughly constant. Karpathy's framing.
- **H-LAM/T** — Engelbart 1962: Human using Language, Artifacts, Methodology, Training. Current rollouts ship the artifact; methodology and training lag.
- **HAIJCS** — Human-AI Joint Cognitive System (Xu & Gao 2024). Bridge from CSE to LLM teaming.
- **ironies of automation** — Bainbridge 1983: AI handling routine work degrades human capacity to catch the rare critical errors.
- **jagged frontier** — Dell'Acqua 2023: AI capability is heterogeneous across sub-tasks; using AI inside the frontier helps, outside causes harm.
- **L2D** — Learning to Defer; Madras / Mozannar formal allocation theory.
- **metacognitive demands** — Tankelevitch 2024 best-paper framework: GenAI reduces production load but increases planning, evaluation, and calibration load.
- **MABA-MABA** — "Men Are Better At, Machines Are Better At." Dekker & Woods 2002 critique of static function allocation.
- **MRT (Multiple Resource Theory)** — Wickens. Cognitive resources are partitioned by processing stage, sensory modality, and processing code; tasks competing for the same resource interfere superlinearly while tasks using different resources can be parallelized cheaply. Predicts tool-stack attention overload (Cursor + ChatGPT + meeting incurs cost beyond the sum). Lit review notes MRT is "absent from AI workflow research."
- **read/write asymmetry** — Cognition: multi-agent works for read-heavy tasks but breaks for write-heavy unless writes are serialized.
- **routing > model selection** — Paterson 2026: across 38 real daily tasks and 15 models, dispatching by task type beats picking a single best model.
- **self-automator** — Randazzo 2026 third Mollick-typology mode: full delegation with periodic oversight.
- **sycophancy in feedback loops** — Randazzo HBS 26-021: AI escalates persuasive justification on pushback.
- **Tools for Thought** — Microsoft Research program on AI-augmented knowledge work.
- **verification cost** — the cost of checking whether an AI output is correct. Vasconcelos 2023 reframing: overreliance is rational when verification is expensive relative to expected payoff.

---

## Writeup
*topic: technology-utilization-architecture · stage: writeup · pass 0 · not-started*

Stub. The long-form synthesis is produced after all five preceding stages are complete.

Coming soon.

---

## Build
*topic: human-psych-variation · stage: build · pass 6 · complete*

A reader's tool for the psychology of individual differences. Pick a trait, see the three plain-language buckets (direct genes / family setup / environment + chance) instead of the V(A_d)/V(A_LD)/V(A_i) decomposition. Plus the four motivated-reasoning traps the field gets caught in, the asymmetric environmental-effects finding, three "heritability ≠ destiny" misreadings, and a seven-bullet take-away. Translates the formalization and data pipeline into something a non-specialist can actually use.

## TLDR

The model formalization produced one equation per person and seven variance components. The data pipeline produced eight tested predictions and seven downloadable CSVs. Both are correct, both are useful for someone who already speaks the vocabulary, and neither does what the topic statement asked: produce something useful for someone who wants to understand how and why people differ without being captured by motivated reasoning from any direction.

This build is that translation layer. It collapses the seven-component variance decomposition into three plain-language buckets — *direct genes*, *family setup*, *environment + chance* — picks the ten traits a reader most likely cares about, and for each one shows the bucket breakdown, the key environmental levers (when relevant), and the two specific ways the most common political readings of that trait go wrong. Plus four secondary views: the four motivated-reasoning traps the field gets caught in (with what each side cites correctly and ignores), the asymmetric environmental-effects finding (severe insults cost 10–30 IQ points; enrichment above normal yields a few at most — the single most useful action-oriented insight), three "heritability ≠ destiny" misreadings with worked examples, and a seven-bullet take-away that holds up across mainstream behavior genetics in 2026.

If you want to engage with the math, the [model stage](/ai-research/human-psych-variation/model) has the parametric dashboard and the [data stage](/ai-research/human-psych-variation/data) has the prediction-by-prediction empirical tests. This page is for the reader who wants to come away knowing what to actually believe.

<PsychVariationExplorer client:load />

## How to use this

The default view is **trait lookup**. Pick a trait — adult cognitive ability, schizophrenia, height, political orientation — and see the three-bucket breakdown plus the trait-specific traps and take-away. Most readers should start there, then move through the four secondary views in order.

A few framing notes:

The **three buckets are not orthogonal categories of cause**. They are three plain-language groupings of the seven model-formalization variance terms (`A_d`, `A_LD`, `A_i`, `C`, `E_m`, `E_s`, `I`). Direct genes is the within-family direct-effect slice — the part that is unambiguously direct biological causation. Family setup combines AM-induced LD, genetic nurture, and residual shared environment — all the things that get counted as "genetic" in twin studies but are not direct biological causation. Environment + chance combines measured non-shared environment and stochastic developmental noise. The split is pedagogical; the [model](/ai-research/human-psych-variation/model) shows the underlying seven-term decomposition.

The **four-traps view is opinionated** in a way the other views are not. The label "trap" assumes that motivated reasoning is what produces these positions, which is not entirely fair — most people cite the evidence they have seen and have not personally vetted what they have not seen. The integrated reading at the bottom of each trap card is the closest the artifact comes to a normative claim about how the field should be read, and it is not algorithmically derivable from the data alone. If you disagree with one of the integrated readings, the [topology stage](/ai-research/human-psych-variation/topology) has the underlying graph.

The **asymmetry finding** is the most action-relevant single insight in the topic. If you only take one thing away from this work, take that one — the population-level cognitive levers run almost entirely through preventing severe insults, not through optimizing within normal. Most parental anxiety and policy expenditure on enrichment is misallocated relative to where the empirical effect sizes are.

The **seven take-aways** are calibrated to be the things a behavior-geneticist in 2026 would actually defend in a public talk. Finer-grained claims (specific magnitudes per trait, mechanism per finding, what polygenic scores measure causally) sit downstream of these and are more contested.

## What this is not

It is not a prediction tool. There is no model that takes your demographics, your parents' phenotypes, or your DNA and outputs a predicted trait value. The science does not currently support that for psychological traits, and the [data stage](/ai-research/human-psych-variation/data#h5--pgs-portability-decay-supported) shows why — polygenic scores trained on European-ancestry data lose 30–80% of their accuracy across other ancestries, and within-family direct effects are often less than half of population-level prediction.

It is not policy advice. The asymmetry finding has clear implications for cognitive intervention (lead remediation has higher effect-per-dollar than enrichment programs), but turning empirical asymmetries into policy involves trade-offs the science does not adjudicate.

It is not a complete picture. Three open questions named in the [model stage](/ai-research/human-psych-variation/model#8-open-questions-that-the-model-exposes-stage-4-inputs) (the Plomin/Turkheimer dispute about what polygenic scores measure, the mechanism behind the Gender Equality Paradox, the magnitude of assortative-mating correction across the full psychiatric cross-disorder rg matrix) are not answered here because the field has not answered them. The honest reading is "we don't know yet"; the artifact does not pretend otherwise.

## Connection to the rest of the pipeline

The trait-lookup numbers are computed directly from `public/data/human-psych-variation/heritability_estimates.csv` (the Stage-4 input), with the H2 partition (`V(A_LD) = m·h²`) and the genetic-nurture identity (`V(A_i) = (β_i/β_d)² · V(A_d)`) applied as in the [model formalization §3.3 pass 5](/ai-research/human-psych-variation/model). The asymmetry view's exposure list comes from `environmental_effects.csv` (the H7 input). The four-traps view materializes the [topology stage's](/ai-research/human-psych-variation/topology#variant-d-politicization-map--where-does-motivated-reasoning-concentrate) Variant D distortion-to-target matrix (D1–D4) into reader-facing cards.

A future stretch would promote some of this to `/dashboards/human-psych-variation` as a public dashboard that lets the visitor enter their own per-trait estimates and see the buckets recompute. That is one of the planned dashboard slots in the site PRD but is out of scope for the first build.

---

## Data
*topic: human-psych-variation · stage: data · pass 9 · complete*

Empirical pipeline that confronts the model's eight testable predictions with currently-published consortium estimates. Seven hold cleanly (AM partition, Wilson curve, multivariate-D gap, PGS portability decay, xAM inflation, environmental causes, G×E interaction-conditional); one (the cross-paper method gradient) is mixed in an informative way. Curated CSVs (downloadable) + Python pipeline + interactive findings panel.

## TLDR

This stage takes the model's eight concrete predictions about how human psychological variation breaks down — how much of trait-variance is genetic-direct vs. genetic-via-parents vs. assortative-mating-induced vs. measured-environment vs. gene-environment-interaction — and confronts each one with currently-published consortium numbers. Seven predictions hold cleanly. One — that the four standard heritability estimators (twin, whole-genome-sequence, common-SNP, within-family) should line up in a strict numeric ordering — is mixed across published papers because each paper uses different cohorts and methods, but holds within any single paper that runs the comparison properly. That "mixed" verdict turns out to be informative rather than a model failure: it tells you the cross-paper landscape is noisier than a literal subtraction of estimates suggests.

Headline empirical findings: assortative mating (people pairing with partners of similar traits) creates linkage between trait-relevant alleles, contributing a Crow-Felsenstein V(A_LD)/V(A_AM) share of ~20% for height, ~22% for educational attainment, ~36% for schizophrenia, ~33% for ADHD, ~36% for autism (the AM-strong psychiatric block; affective disorders sit lower at ~6–14%). These percentages are population-level decompositions of V(A) at AM equilibrium — *not* "fraction of twin h² explained by AM." Falconer's classical twin formula is itself biased downward by AM, and for socially-structured traits the empirical gap between twin h² and within-family h² is dominated by genetic nurture and equal-environments-assumption violations, not AM-induced LD (see §2 H2 caveats for the corrected interpretation). Heritability of cognitive ability rises from ~20% in early childhood to ~80% in adulthood along a logistic curve fitted to Bouchard 2013's seven anchor points within 1.8 percentage points. Multivariate sex-difference effect sizes are large (16PF Mahalanobis distance D = 2.7) when computed at the latent-variable level with measurement-error disattenuation, but only D ≈ 1 at the raw observed level — the entire "Mars-and-Venus" framing trap lives inside that disattenuation correction, not inside the multivariate algebra. Polygenic scores trained on European-ancestry data lose ~37%, ~50%, and ~78% of their accuracy in South Asian, East Asian, and African ancestry samples respectively (Martin 2019), consistent with Ding 2023's independent continuous-distance result of Pearson r = −0.95 across 84 traits. Cross-trait assortative mating accounts for ~74% of the variance in reported psychiatric cross-disorder genetic correlations (Border 2022, 132 trait pairs). The small set of measured environments with replicated causal effects on cognition is asymmetric: severe insults (lead, fetal alcohol, deprivation, malnutrition) cost 10–30 IQ points, while enrichment above normal yields at most a few points per intervention. And gene-by-environment interaction (V(I)) shows the classic Scarr-Rowe pattern of higher heritability at higher SES *only in US samples* (Tucker-Drob & Bates 2016 meta-analysis: a' = 0.074, p &lt; .0005); equity-buffered W. European / Australian samples show no such interaction (a' = −0.027, n.s.) — the cross-national heterogeneity is exactly what the model predicts under "V(I) is small at typical environmental variance, larger at extreme tails."

The pipeline is intentionally small. Seven curated CSVs (one per data type, every cell source-cited), a single ~350-line Python script that produces every chart on this page, dependencies pandas + numpy + scipy. **Inputs are downloadable** from [/data/human-psych-variation/](/data/human-psych-variation/). Stage 5 (build) consumes the CSVs directly. What the pipeline does **not** answer: whether polygenic scores measure direct biological causation or correlated environments (the Plomin–Turkheimer dispute, undecidable without a within-family environmental intervention no group has run); the mechanism behind the Gender Equality Paradox (needs cross-society multivariate panels that don't exist at scale); and the full assortative-mating-corrected psychiatric genetic-correlation matrix (active research, not yet pipeline-runnable from public summary statistics).

## A few terms

The data stage inherits the model formalization's vocabulary. If you arrived here without reading [the model stage](/ai-research/human-psych-variation/model), the terms below cover what's used in the prose:

- **Heritability (h²).** The fraction of variance in a trait, across people in a population, that tracks genetic differences. A *population* statistic, not an individual one — saying "IQ is 70% heritable" does not mean 70% of any one person's IQ is genetic.
- **Twin h², SNP h², WGS h², within-family h².** Four ways to estimate heritability, each picking up a slightly different slice of the underlying genetic variance. Twin: from MZ vs. DZ similarity. SNP: from GWAS effect sizes on common variants only. WGS: SNP plus rare variants. Within-family: from sibling differences, controls for parental environment.
- **Assortative mating (m).** The correlation between partners on a trait — partners are similar on educational attainment (m = 0.55), height (m = 0.24), political views (m = 0.58). The model's claim is that AM creates linkage between causal genetic variants, inflating measured h² by a calculable amount.
- **Polygenic score (PGS).** A weighted sum of risk alleles per person, used to predict the trait. PGS R² is the variance the score explains in a held-out sample.
- **Mahalanobis D.** The multivariate analogue of Cohen's d for sex (or any group) differences across multiple correlated measurements.
- **V(E_m).** The model's variance bucket for *measured* non-shared environment — exposures with named causal coefficients (lead, schooling, iodine, etc.).
- **V(I).** The model's variance bucket for *interaction* effects: gene × environment, gene × gene (epistasis), gene × age. The model's specific claim is V(I) is small at typical PGS-by-environment scale but larger when environmental variance includes extreme tails — tested in H8 below.
- **Scarr-Rowe interaction.** The hypothesis (founded in Turkheimer 2003's US data) that IQ heritability is lower in low-SES families than in high-SES families. Tucker-Drob & Bates 2016 meta-analyzed it and found the pattern replicates in US samples but vanishes in W. European / Australian samples. The cross-national heterogeneity is the H8 test of V(I).

<PsychVariationData client:load />

## How to read this stage

The panel above is the artifact. The prose below is the spec.

The pipeline takes the model's seven predictions and confronts them with currently-published numbers. Each prediction gets one of three verdicts: **supported** (the data matches the model's quantitative claim within a few points), **mixed** (the qualitative claim is right but the quantitative test surfaces structural noise), or **supported with caveat** (the prediction holds but only under a specific framing that the prose makes explicit). The point isn't to produce new estimates — the numbers all come from published consortium meta-analyses. The point is to align them in one place so the model's predictions can be tested cleanly, and to flag where the literature is good enough vs. where the field hasn't yet collected what the model would need.

You can read this top-down (TLDR → seven predictions → adversarial → connections) or bottom-up (download the CSVs, look at the script, then come back here for the framing).

## 1. Pipeline architecture

Seven curated CSVs in `public/data/human-psych-variation/` (downloadable from the live site, tracked in git):

| File | Rows | Purpose |
|---|---|---|
| `heritability_estimates.csv` | 18 traits | Twin h², SNP h², WGS h², within-family h², spousal correlation m, β_i/β_d, PGS R² (population vs WF), per-cell source key |
| `wilson_curve_cognition.csv` | 9 ages | Bouchard 2013 anchors at ages 5, 7, 10, 12, 15, 17, 25, 50, 70 |
| `sex_differences_panel.csv` | 7 panels | Per-panel univariate d̄, ρ̄, n_dimensions, observed D, disattenuated D — Hyde 2008, Su 2009, Schmitt 2008, Del Giudice 2012, Kaiser 2020, Ritchie 2018 |
| `pgs_portability.csv` | 13 rows | PGS R² ratio (relative to European training) by target ancestry × trait, with genetic distance |
| `environmental_effects.csv` | 10 exposures | Per-exposure causal effect sizes on cognition: lead, schooling, iodine, FAS, PM2.5, deprivation, malnutrition, breastfeeding, adoption, parenting |
| `gxe_interactions.csv` | 7 rows | Tucker-Drob & Bates 2016 meta-analysis a' by region (US vs non-US), Turkheimer 2003 anchors, German replication |
| `sources.csv` | 23 papers | Full citation, DOI/URL, what each paper is used for |

A single Python script (`pipeline.py`) reads the inputs, computes derived quantities (AM partition, Wilson logistic fit, equicorrelated D, PGS portability slope, genetic-nurture variance contribution, environmental-effect summary), and writes:

- `out/method_gradient.csv` — per-trait alignment with deltas
- `out/am_partition.csv` — `r_δ`, V(A_d), V(A_LD) per trait
- `out/genetic_nurture.csv` — V(A_i) and cross-term per trait
- `out/sex_diff.csv` — equicorrelated D per panel
- `out/findings.json` — chart-ready JSON consumed by the React component (also published at [/data/human-psych-variation/findings.json](/data/human-psych-variation/findings.json))
- `out/findings_table.md` — markdown audit table of the seven predictions

Dependencies: `pandas`, `numpy`, `scipy`. No web fetches, no external services, no individual-level genetic data. Reproduces in under 1 second on a laptop.

## 2. Seven predictions, seven tests

### H1 — Method gradient (mixed)

**Claim.** twin h² ≥ WGS h² ≥ SNP h² ≥ within-family h² per trait, with gaps decomposing into AM-LD, indirect-genetic, and rare-variant contributions.

**Result.** Across 15 traits with at least two published estimators, the strict ordering holds for 9 (all 2-estimator rows where twin h² > SNP h²) and fails for 6 (all 3-estimator rows). Every failure is the same: SNP h² is *lower* than within-family h² for socially-stratified traits — for height, SNP h²=0.50 vs. within-sibship h²=0.78; for EA, SNP h²=0.13 vs. within-sibship h²=0.15; for IQ adult, SNP h²=0.20 vs. extrapolated WF h²=0.50. This is not a model failure but a structural property of LDSC: it captures common-variant additive variance in unrelated populations and undercounts the rare-variant share, while within-family designs capture rare variants implicitly through transmission. The model's V(A_d) is naturally higher than what SNP h² estimates.

**Within a single paper, the prediction holds cleanly.** Howe 2022 (N=178,086 siblings) is the only published study that runs population vs. within-sibship GWAS on the same sample. Their Figure 4 shows population effects exceed within-sibship effects for height, EA, age at first birth, # children, cognitive ability, depressive symptoms, smoking — exactly the seven traits the model singles out as having non-trivial indirect-genetic contributions.

**What this teaches.** "twin h² > within-family h²" is the canonical robust finding (always holds). "SNP h² between twin and within-family" is a methodological artifact when applied across papers — the right cross-check is twin vs. within-family directly, leaving SNP h² as a third estimator that answers a slightly different question (common-variant only).

### H2 — AM partition (supported)

**Claim.** `V(A_LD) = m·h²` with the AM equilibrium reached.

**Result.** Predicted V(A_LD) shares of observed h²: educational attainment **22%**, height **20%**, BMI **12%**, schizophrenia **36%**, ADHD **33%**, autism **36%**, bipolar **14%**, MDD **6%**, IQ adult **35%**. Height matches Yengo 2018's reported empirical 14–23% range; EA matches Border 2022's qualitative "substantial fraction" finding.

**The psychiatric numbers were corrected in pass 4.** Pass-1/2/3 used m=0.30 for schizophrenia, ADHD, and autism (cited as "Nordsletten 2016 imputed" without verified value). [Nordsletten 2016](https://pubmed.ncbi.nlm.nih.gov/26913486/) actually reports tetrachoric spousal correlations *greater than 0.40* for all three disorders — moving these from m=0.30 to m=0.45 lifts their predicted V(A_LD) share from ~24% to ~36% of h². This is a real and substantively different reading: about one third of the additive genetic variance for severe psychiatric conditions is *structural assortative-mating-induced LD* rather than independent direct biological signal. The model's prediction stands; the data is more dramatic than pass-1 numbers showed.

**Caveats.** The Crow–Felsenstein partition assumes AM equilibrium. For traits under rapid assortment shifts (EA post-1970), this is approximate. The IQ adult prediction (35%) sits at the upper end and may overshoot — Horwitz 2023's IQ partner correlation r=0.44 comes from a small (N=5,672) meta-analytic sample. For psychiatric disorders, "spousal correlation" is a tetrachoric across a binary diagnosis, which behaves differently than a continuous-trait partner correlation under the same equilibrium assumption — the prediction is qualitatively right but quantitative precision is lower.

**A reviewer correction added in pass 7.** The framing "structural assortative-mating-induced LD" implied that AM is the source of the gap between Falconer twin h² and within-family h² for socially-structured traits. This is incorrect: Falconer's `2·(rMZ − rDZ)` is itself biased *downward* by AM (under positive AM, fraternal twins share more than 50% of trait-relevant alleles, raising rDZ relative to rMZ). The empirical gap between twin h² and within-family h² for socially-structured traits is dominated by genetic nurture and equal-environments-assumption violations, *partially offset* by AM's downward bias on Falconer. The formula `V(A_LD) = m·h²` is mathematically valid as a Crow-Felsenstein population-level decomposition of V(A) at AM equilibrium — Yengo 2018's empirical 14–23% V(A_LD)/V(A) for height matches the formula prediction at the population level — but it does NOT predict the twin-vs-within-family gap, and the percentages reported above ("22% of h² for EA" etc.) should be read as population-level V(A_LD)/V(A_AM) shares, not as "fraction of twin h² explained by AM." The cross-trait AM result (Border 2022, H6 below) is independent of this issue and stands as reported.

### H3 — Wilson logistic curve (supported)

**Claim.** `h²(t) = h²_∞ / (1 + exp(−k·(t − t₅₀)))` for cognitive ability across age.

**Result.** Fitted to Bouchard 2013 anchors:

```
h²_∞ = 0.81
t_50 = 9.0 years
k    = 0.27 / year
```

Max residual: 1.8 percentage points (at age 12). The earlier saturating-exponential form (Stage 3 pass 2) had max residual 32 pp at age 5. The logistic is the smallest functional change that matches the empirical sigmoidal pattern, and the fitted parameters are within sampling noise of the model's prior values (h²_∞=0.80, t_50=9.0, k=0.30).

### H4 — Equicorrelated D vs disattenuated D (supported with caveat)

**Claim.** Equicorrelated `D² = d̄²·n / (1 + (n−1)·ρ̄)` is a pedagogical anchor; the gap to disattenuated D is exactly the latent-variable correction.

**Result.** For Del Giudice 2012's 16PF panel (n=15, d̄=0.50, ρ̄=0.18): equicorrelated D = 1.03; disattenuated D = 2.71. Ratio: 2.6×. The equicorrelated approximation is *quantitatively wrong* for high-dimensional disattenuated panels — but not because of an algebra error. The 2.6× factor is the disattenuation correction: latent-variable modeling magnifies effect sizes by ~1/√reliability per factor before aggregation.

**For the public-discourse framing trap (univariate d small vs. multivariate D large), this means:** the gap exists at *both* observed and latent levels (D=1.03 vs d=0.05 is already a 20× scale-up). Disattenuation pushes it further. Both Hyde 2005 ("similarities hypothesis") and Del Giudice 2012 ("Mars and Venus") are correct about their respective objects of measurement.

### H5 — PGS portability decay (supported)

**Claim.** PGS accuracy decays with genetic distance from the training population.

**Result.** [Ding et al. 2023](https://www.nature.com/articles/s41586-023-06079-4) reports Pearson r = −0.95 between continuous PCA-based genetic distance and PGS R² across 84 traits (their analysis on individual-level UK Biobank + ATLAS data, N≈524k, which we don't have access to). Independent categorical-ancestry estimates corroborate the trend: [Martin et al. 2019](https://www.nature.com/articles/s41588-019-0379-x) reports relative-accuracy reductions of 37%, 50%, and 78% in South Asian, East Asian, and African ancestries vs. European training; per-trait, [Okbay 2022 EA4](https://www.nature.com/articles/s41588-022-01016-z) reports near-zero EA-PGS accuracy in African samples; [Yengo 2022](https://www.nature.com/articles/s41586-022-05275-y) reports height-PGS accuracy at 10–20% of European levels in non-European ancestries; [Trubetskoy 2022](https://www.nature.com/articles/s41586-022-04434-5) reports schizophrenia-PGS accuracy at ~30% in African samples. The pipeline aggregates these per-ancestry literature anchors into one panel and computes a slope as a sanity check that the literature is internally consistent (Pearson r = −0.99 on 11 anchored rows). **This is not an independent replication of Ding 2023** — those rows are themselves drawn from primary papers — but it is a defensible visualization of the convergent empirical pattern.

**Why this matters for the L4 firewall.** The model's between-population scope restriction is structurally argued: there is no μ_pop term in the generating function. The empirical evidence for *why* the restriction matters is the portability decay — the same SNP "effect sizes" do not estimate the same causal coefficients in different populations. Causal architecture is not portable; descriptive variance partitions arguably are, but not for cross-population mean comparisons.

### H6 — Cross-trait AM inflation (supported)

**Claim.** Cross-trait assortative mating accounts for a substantial fraction of reported psychiatric cross-disorder genetic correlations.

**Result.** Border 2022 (UK Biobank N=40,697 spousal pairs, 132 trait pairs): R² = 0.7432 (95% CI: 0.67–0.82) between phenotypic cross-mate correlations and reported genetic correlations. Across 6 psychiatric disorders × 5 generations: average xAM share γ̂ = 0.29. Anxiety × MDD: γ̂ = 0.21 (95% CI: 0.17–0.25). AUD × schizophrenia: γ̂ = 0.83 (95% CI: 0.59–1.24).

**Interpreting γ̂.** The γ̂ statistic is the ratio of the xAM-alone-projected genetic correlation to the empirical genetic correlation. A value near 1 is *consistent with* xAM accounting for the entire reported rg — it does not *prove* xAM is the cause, since alternative causal architectures (genuinely shared biology with the same effect-size profile) could produce the same ratio. But γ̂ values bounded well below 1 require an additional shared-biology contribution beyond what xAM alone can explain. The Border result is therefore a pressure-test: if reported cross-disorder rg estimates were entirely about shared biology, γ̂ would be small; the average γ̂ = 0.29 with significant pair-level variance shows the literature's cross-disorder rg estimates carry an xAM contribution that is empirically non-trivial and pair-specific.

**Implication.** The within-trait V(A_LD) term is the within-trait analogue of cross-trait xAM. Same operation (LD created by non-random mating among causal alleles); they show up in different summary statistics.

### H7 — Environmental causes (supported)

**Claim.** The model's V(E_m) term — variance contribution of measured non-shared environment — is non-empty: a small set of exposures have large, replicated, causal effects on cognitive outcomes.

**Result.** Per-exposure effect sizes:

| Exposure | Effect on IQ | Source | Design |
|---|---|---|---|
| Schooling, per year | +1 to +5 pts (mean +3.4) | [Ritchie & Tucker-Drob 2018](https://journals.sagepub.com/doi/10.1177/0956797618774253) (600k participants, 3 designs) | Quasi-experimental meta |
| Breastfeeding (PROBIT RCT) | +3.2 pts | Kramer 2008 (N=17,046) | Cluster RCT |
| Within-Western-normal parenting | ~0 to +1 pts | Plomin & Daniels 1987 meta | Within-family twin |
| PM₂.₅, per 1 µg/m³ | −0.27 pts | Aghaei 2024 meta | Observational meta |
| Lead, blood 1→10 µg/dL | **−6.2 pts** (CI −8.6 to −3.8) | [Lanphear 2005](https://pmc.ncbi.nlm.nih.gov/articles/PMC1257652/) (N=1,333, 7 cohorts) | Pooled longitudinal |
| Iodine, severe deficiency | **−10 pts** (recovers +8.7 with supplementation) | Bougma 2013 | Observational + RCT |
| Adoption: high → low SES | **−12 pts** | Capron & Duyme 1996 (N=38) | Natural experiment |
| Severe psychosocial deprivation | **−15 pts** | Nelson 2007 BEIP (N=136) | Natural experiment |
| Severe chronic malnutrition | **−15 pts** | Grantham-McGregor 2007 | Observational |
| Prenatal alcohol (full FAS) | **−30 pts** | Streissguth 2004 | Observational + MR |

**Asymmetry is the headline finding.** Removing severe insults (lead, malnutrition, deprivation, FAS) recovers double-digit IQ points; enrichment above normal (better parenting, breastfeeding) yields single-digit gains at most. The variance-share interpretation V(E_m)/V(P) depends on each exposure's prevalence in a given population — sparse-but-large exposures (FAS, severe deprivation) contribute little to population variance despite large per-person effects, while moderate-but-common exposures (variable schooling quality, low-grade lead) contribute more. This is why the high-h² findings of behavior genetics coexist with large environmental effects without contradiction: heritability is a population-variance statistic, individual environmental effects can be enormous, and most populations have already removed the worst tails.

### H8 — G×E interaction (V(I) bucket) — supported conditional

**Claim.** The model's V(I) term — variance contribution of gene-environment interaction — is small at typical PGS-by-environment scale but larger when environmental variance is wide enough to include extreme tails.

**Result.** [Tucker-Drob & Bates 2016](https://pmc.ncbi.nlm.nih.gov/articles/PMC4749462/) meta-analyzed 43 effect sizes across 14 independent studies (24,926 twin / sibling pairs, ≈50,000 individuals) testing the Scarr-Rowe Gene × SES interaction on intelligence. Their Purcell-biometric-model coefficient `a'` represents the expected change in the additive genetic regression on intelligence per SD of SES. Reported numbers:

| Sample | a' | SE | Significance | N pairs |
|---|---|---|---|---|
| US-pooled | **+0.074** | 0.020 | p < 0.0005 | 11,340 |
| Non-US-pooled (W. Europe / Australia) | **−0.027** | 0.022 | p = 0.22 (n.s.) | 13,586 |
| Overall pooled | +0.029 | 0.019 | p = 0.14 (n.s.) | 24,926 |

Plus the founding observation from [Turkheimer 2003](https://journals.sagepub.com/doi/10.1111/1467-9280.t01-1-01498): IQ heritability h² ≈ 0.10 in low-SES US families, rising to h² ≈ 0.72 in high-SES US families. And independent null replication in Germany (Spengler 2018: a' = −0.01, n.s.).

**Interpretation.** The cross-national heterogeneity is the empirical confirmation of the model's "extreme-environment-threshold" reading. US samples have wider environmental tails — extreme low-SES exists in larger numbers, with worse low-SES conditions, than in W. European or Australian welfare-state samples. The model predicts V(I) shows up exactly where the low-SES tail is wide enough to include genuine environmental constraint that suppresses genetic expression. Equity-buffered samples truncate that tail; the interaction shrinks toward zero. The verdict is "supported conditional" because the prediction is conditional on environmental variance: the same model that predicts a' ≈ 0.074 in US samples predicts a' ≈ 0 in equity-buffered samples, and both predictions match.

**Caveat.** The Scarr-Rowe finding is itself contested in the literature. Several individual replications have been null even within US samples (e.g., Hanscombe 2012); the pooled US a' = 0.074 is moderate but not large. The model claim "V(I) is small at typical PGS-by-environment scale" is *most* supportable; the stronger claim "G×E reliably appears at extreme tails" is supportable but with wider error bars than H1–H7.

## 3. Headline numbers

| Statistic | Value | Source |
|---|---|---|
| Mean h² across human traits | 0.49 | [Polderman 2015](https://www.nature.com/articles/ng.3285) (17,804 traits, 14.5M twin pairs) |
| Non-transmitted EA-PGS effect | 29.9% of transmitted | [Kong 2018](https://www.science.org/doi/10.1126/science.aan6877) (N=21,637) |
| EA4 within-family direct effect | ~50% of population PGI | [Okbay 2022](https://www.nature.com/articles/s41588-022-01016-z) (N=3M) |
| Height WGS h² | 0.68 (SE 0.10) | [Wainschtein 2022](https://www.nature.com/articles/s41588-021-00997-7) (N=25,465) |
| WGS captures of pedigree h² | 88% | [Wainschtein 2025](https://pubmed.ncbi.nlm.nih.gov/41225014/) (N=347,630, 34 traits) |
| Spousal correlation EA | 0.55 | [Horwitz 2023](https://www.nature.com/articles/s41562-023-01672-z) (N≈1.9M pairs) |
| Spousal correlation political | 0.58 | Horwitz 2023 |
| Spousal correlation IQ | 0.44 | Horwitz 2023 (N=5,672 pairs) |
| Cross-trait AM inflation R² | 0.74 (CI: 0.67–0.82) | [Border 2022](https://www.science.org/doi/10.1126/science.abo2059) (132 pairs) |
| Avg psychiatric γ̂ (xAM share) | 0.29 | Border 2022 |
| Wilson curve h²_∞ (cognition) | 0.81 (fit) | Pipeline fit to [Bouchard 2013](https://pubmed.ncbi.nlm.nih.gov/23919982/) |
| Wilson curve t_50 (cognition) | 9.0 years (fit) | Pipeline fit |
| 16PF Mahalanobis D observed | 1.03 | Equicorrelated approximation |
| 16PF Mahalanobis D disattenuated | 2.71 | [Del Giudice 2012](https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0029265) |
| PGS R² ~ genetic distance | r = −0.95 (continuous) | [Ding 2023](https://www.nature.com/articles/s41586-023-06079-4) (84 traits, 524k indivs) |
| PGS accuracy in AFR vs EUR | 22% relative (78% reduction) | [Martin 2019](https://www.nature.com/articles/s41588-019-0379-x) (across-trait avg) |
| Lead 1→10 µg/dL → IQ | −6.2 pts | [Lanphear 2005](https://pmc.ncbi.nlm.nih.gov/articles/PMC1257652/) |
| Schooling/year → IQ | +1 to +5 pts | [Ritchie & Tucker-Drob 2018](https://journals.sagepub.com/doi/10.1177/0956797618774253) |
| G×SES (US) | a' = +0.074 (p < .0005) | [Tucker-Drob & Bates 2016](https://pmc.ncbi.nlm.nih.gov/articles/PMC4749462/) (43 effects, 25k pairs) |
| G×SES (non-US) | a' = −0.027 (n.s.) | Tucker-Drob & Bates 2016 |
| Turkheimer 2003 IQ h² range | 0.10 (low SES) → 0.72 (high SES) | [Turkheimer 2003](https://journals.sagepub.com/doi/10.1111/1467-9280.t01-1-01498) |

## 4. Analytical choices

The pipeline has six judgment calls. Each is flagged in the script as `# ASSUMPTION:`. The most consequential:

1. **Twin h² as h²_observed for AM partition.** Twin h² is closer to the AM-equilibrium quantity than SNP h². For traits without twin estimates we fall back to SNP h².
2. **AM equilibrium assumption.** The Crow–Felsenstein partition assumes mating regimes are stable. For EA (post-1970 educational expansion) this is approximate.
3. **k ≈ 0.5·m for the genetic-nurture cross-term.** The AM-coupling parameter k is empirically 0.1–0.5 for AM-strong traits; we interpolate.
4. **Equicorrelated Σ for multivariate D.** Real personality covariance matrices have hierarchical structure; the equicorrelated approximation is pedagogical, not quantitative for high-dimensional panels.
5. **PGS portability linear in genetic distance.** Ding 2023 reports a strong linear correlation. For genetic distances near zero the relationship may be non-linear. Our 5-trait curated panel is small.
6. **Within-family h² for IQ extrapolated.** No within-family GWAS h² has been published for cognitive ability at the same scale as Howe 2022's other traits. We extrapolate from EA's WF h² and the EA-IQ rg.

## 5. What the pipeline does *not* deliver

Three open questions from the model's §8 list are *not* sharpened by this stage, despite being framable:

- **O1 — PGS interpretation (Plomin/Turkheimer).** The decisive test is whether within-family β_d moves under environmental intervention. No paper has the design — Sacerdote 2007 Korean adoption comes closest but predates within-family GWAS. **Status:** open.
- **O3 — Gender Equality Paradox.** Tests whether multivariate sex-difference D depends on Σ-by-society in addition to μ-by-society. Stoet & Geary 2018 / Schmitt 2008 give univariate cross-cultural d's; the multivariate piece requires Σ-by-society panels that do not yet exist at scale. **Status:** likely answerable in the next 5 years.
- **O7 — xAM-corrected full psychiatric rg matrix.** Border 2022 establishes the principle on 6 disorders. Applied at scale to the full PGC cross-disorder matrix, the corrected rg's are likely smaller — but no group has done the correction systematically. **Status:** active research.

For these three, the Stage-4 honest answer is "the pipeline frames them but doesn't resolve them."

## 6. Adversarial + steelman

Four objections to the pipeline. The strongest version of each, then the honest response.

### Objection 1 — This is variance bookkeeping, not new analysis

The pipeline arranges other people's published estimates in a table and runs simple closed-form computations on top. It does not produce new heritability estimates, does not analyze raw data, and does not test causal mechanisms. Calling it "an empirical pipeline" overstates what is actually a literature-alignment exercise.

**Steelman.** True at the bookkeeping level. A real empirical pipeline would pull GWAS summary statistics, run LDSC against multiple traits, replicate Howe 2022's within-sibship analysis on UK Biobank data, and compute fresh AM-LD partition estimates per trait. That requires individual-level genetic data we do not have access to and would not be appropriate to ship from a content site.

**Response.** Conceded as a scope restriction. The pipeline's value is at the meta-level: it confronts the model's predictions with the literature that already exists and surfaces what does and does not match. Three contributions are genuinely new even at this scale: (a) per-trait AM-partition predictions computed at the granularity of single traits with current Horwitz 2023 m-values, which Border 2022 / Yengo 2018 framed only at the single-trait level; (b) the equicorrelated-D vs. disattenuated-D bridge that locates the entire Hyde-vs-Del-Giudice gap quantitatively in the disattenuation correction; (c) the explicit reframing of H1 as "within-paper holds, cross-paper noisy" with the structural reason. None of these required new data analysis, but none were available in one place before.

### Objection 2 — The CSV is too small to support strong claims

18 traits is a small panel. The headline-sounding patterns (e.g., "the AM partition holds across AM-strong traits") rest on roughly six traits. A bigger panel might tell a different story.

**Steelman.** True for any single trait — the AM partition prediction for IQ adult lands at the upper end of the empirical range and could be wrong. For the multivariate-D module, only one panel (16PF Del Giudice) drives the pedagogical claim; the same algebra on a different instrument might give a smaller disattenuation gap.

**Response.** The headline patterns are robust within the curated traits and consistent with primary-literature meta-analyses (Polderman 17,804 traits, Border 132 pairs, Horwitz 22 traits + 133-trait UK Biobank scan). Adding another 50 traits would not change the qualitative result for H2 or H6 because those rest on consortium meta-analyses not single-CSV cells. The single-CSV results are calibration checks, not new estimation. Where the pipeline does need more data — H5 portability with 13 hand-curated rows — this is flagged explicitly as Objection 4 below.

### Objection 3 — Border 2022 is a single high-profile paper with significant methodological pushback

Resting H6 on a single 2022 paper from one group is fragile. xAM as a confounder of psychiatric cross-disorder rg has been proposed by other authors (Howe 2024, Cai 2025 commentary) but Border's specific R²=0.74 figure and the 5-generation-equilibrium assumption it depends on have been pushed back on. The "γ̂ averages 0.29" claim depends on a specific xAM dynamics model.

**Steelman.** Conceded. R²=0.74 may shrink under different equilibrium assumptions; γ̂ values for specific pairs may move under alternative AM models. The aggressive interpretation ("xAM accounts for ~30% of psychiatric rg") is doing motivated work in the discourse and would benefit from independent replication by groups outside the Border / Keller cluster.

**Response.** The model's H6 prediction does not depend on Border's specific γ̂ values — it depends on the qualitative claim that cross-trait AM affects rg estimates non-trivially. That qualitative claim has independent support: Howe 2022's within-sibship estimates of EA-BMI rg attenuate to near-zero, Yengo 2018 establishes within-trait AM-LD inflation for height, and the within-trait V(A_LD) prediction (H2) is tested independently from any cross-trait psychiatric finding. The data.mdx prose treats Border 2022 as suggestive about the magnitude rather than dispositive. This was strengthened in pass 2 — the γ̂ wording is now "consistent with xAM accounting for X%" rather than "X% caused by xAM."

### Objection 4 — H5 PGS portability is circular as a test

Pass 1 framed H5 as "replicating Ding 2023's r = −0.95 on a curated 5-trait panel and getting r = −0.98." That was circular: the curated rows were themselves rough approximations of Ding's continuous-distance pattern, so the resulting slope was internal to the curation, not an independent test.

**Response (pass 3 fix).** The CSV was refactored to use named per-ancestry literature anchors instead — Martin 2019 across-trait averages (37%/50%/78% accuracy reduction in SAS/EAS/AFR vs. EUR), Okbay 2022 EA in AFR (relative R² ~10%), Yengo 2022 height in AFR (~20%), Trubetskoy 2022 SCZ in AFR (~30%). The pipeline still computes a Pearson r on this aggregated panel (now r = −0.99), but the prose now describes it honestly as "internally consistent literature-anchored trend, consistent with Ding 2023's independent continuous-distance result," not as a replication. The strong empirical claim — that PGS accuracy collapses across ancestry distance — rests on Ding 2023's primary analysis, with Martin 2019 / Okbay 2022 / Yengo 2022 / Trubetskoy 2022 as independent corroboration on different cohorts and methods.

## 7. Connection to model cruxes

Three of the model's five cruxes (§12) are partly tested by the pipeline:

- **C1 (within-family GWAS unbiased)** — relied upon throughout. Consistent with within-paper agreement across Howe 2022, Okbay 2022, Kong 2018.
- **C2 (AM partition formula)** — partly tested by H2; predictions match Border 2022 / Yengo 2018 within a few points across AM-strong traits.
- **C5 (equicorrelated Σ as useful approximation)** — partly tested by H4; equicorrelated undershoots disattenuated D by 2.6× for the 16PF panel. Crux holds *pedagogically* but not *quantitatively* at high n — same caveat the model already flags.

Cruxes C3 (hyperpolygenic architecture) and C4 (joint identifiability of A_d/A_i/A_LD) are not tested by the pipeline.

## 8. Connections to other work

**To the model dashboard ([/ai-research/human-psych-variation/model](/ai-research/human-psych-variation/model)).** The dashboard's default parameters were set by the model formalization's priors. Several should be updated from the data stage's anchors: spousal correlations for cognitive (`m=0.40` → keep, Horwitz IQ=0.44 confirms), personality (`m=0.15` → keep, Horwitz neuroticism=0.11 close), psychopathology (`m=0.20` → upward to 0.30 for SCZ specifically). The Wilson logistic parameters in the dashboard already match the data-stage fit (h²_∞=0.80 vs. fitted 0.81, t_50=9 exact, k=0.30 vs. 0.27); the tiny discrepancy can either drift the dashboard to the fitted values or note it explicitly.

**To the planned parent-to-child transmission topic.** The V(A_i) data here directly feeds that topic. Howe 2022's within-sibship analysis is the canonical empirical anchor for indirect genetic effects across the seven traits the model singles out (height, EA, age at first birth, # children, cognitive ability, depressive symptoms, smoking). The Kong 2018 non-transmitted-PGS finding (29.9% of transmitted for EA) and the Okbay 2022 EA4 within-family attenuation (~50% of population PGI) are the two anchor numbers the parent-to-child topic should adopt as starting input.

**To the planned evolution-modernity-mismatch topic.** The Wilson curve fit here is the developmental-age analogue of generation-scale changes the mismatch topic will need to address. Pietschnig 2024's finding that the positive manifold itself may be weakening across recent cohorts implies μ(t) is not a one-dimensional trajectory but a moving structure of which abilities are gaining or losing. The data stage's logistic captures developmental motion within a single cohort; the mismatch topic will need to extend it to cross-cohort drift.

## 9. Stage-5 handoff

The Stage-5 build artifact should be a public-facing tool that:

1. Lets a visitor pick a trait and see the per-trait variance decomposition (twin h², SNP h², WGS h², WF h², m, V(A_LD), V(A_i), and the relevant V(E_m) exposures) in a single panel.
2. Surfaces the H1 mixed result honestly: within-paper Howe 2022 chart vs cross-paper alignment.
3. Implements the Mahalanobis-D module with the disattenuation toggle so users can see the framing trap directly.
4. Shows the environmental-effects table with prevalence-weighted variance-share estimates per population (this is the stage-5-specific extension — none of the existing tools do this).
5. Cites a source for every number with a link to the relevant paper.

Inputs are at [/data/human-psych-variation/](/data/human-psych-variation/). Stage 5 can either re-run `pipeline.py` at site-build time or freeze `findings.json` as a static asset.

## 10. Pipeline cruxes

The model stage's §12 listed five load-bearing assumptions of the *formalization*. The pipeline has its own load-bearing assumptions — places where if the assumption fails, specific findings have to be rebuilt. Five matter most.

| Crux | Load-bearing claim | What would flip it |
|---|---|---|
| **D1** | The published estimates I'm citing are correctly extracted from primary sources. ~12 of the highest-uncertainty values were web-verified directly from the cited paper or a PubMed Central mirror; the rest rest on training-time recall plus the cited paper's existence. | A spot-check of the curated CSV against the supplementary tables of any individual paper finds a meaningful discrepancy (>1 SE on the cited estimate). Most of the H2/H3/H6 verdicts would shift correspondingly. |
| **D2** | Twin h² is a usable proxy for h²_observed in the AM partition. The Crow–Felsenstein formula `V(A_LD) = m·h²` assumes h² is the AM-equilibrium quantity; twin h² is the closest readily-available estimate. | A demonstration that twin h² systematically over- or under-estimates the AM-equilibrium h² for the trait class (e.g., if classical ACE leakage from V(A_i) into A is consistently 5+ percentage points). The H2 partition shares would all shift by a similar fraction. |
| **D3** | The equicorrelated approximation captures the *qualitative* multivariate-D framing trap. The pedagogical claim is "stacking weakly-correlated dimensions makes D grow with √n;" the quantitative claim at high-dimensional disattenuated panels is acknowledged not to hold. | A demonstration that real personality covariance matrices have block-structured Σ such that even the qualitative claim fails for the public-discourse-relevant case (16PF / Big Five). H4 would need a worked-example refit using a non-equicorrelated Σ. |
| **D4** | Cross-paper alignment of estimators (twin/SNP/WGS/WF) is structurally noisy enough that within-paper tests are required for clean inference. This is the framing for H1's "mixed" verdict. | A within-paper study that runs all four estimators on the same sample and finds the strict ordering fails. To my knowledge no such study exists; if one publishes and the ordering breaks, H1's "mixed-but-informative" reading collapses to "wrong." |
| **D5** | Per-ancestry PGS-portability anchors from Martin 2019 / Okbay 2022 / Yengo 2022 / Trubetskoy 2022 are concordant with Ding 2023's continuous-distance result. Without individual-level data we cannot compute the continuous-distance slope ourselves; we are taking concordance on faith. | A reanalysis of the cited papers' public summary statistics that finds substantially different per-ancestry decay rates than the headline reports. H5's "consistent with Ding 2023" framing would weaken to "qualitatively matches but quantitatively in dispute." |

The most consequential is **D1** — every other crux assumes the underlying CSV cells are correct. The web-verification round in pass 1 reduced this risk for the dozen highest-stakes numbers; the rest is a calibrated bet on training-time recall and would benefit from a future pass that audits each cell against its primary source.

---

## Lit Review
*topic: human-psych-variation · stage: lit-review · pass 0 · complete*

What the science actually shows about psychological variation — heritability, environmental shaping, gene-environment interaction, sex differences, cognitive capacity. The post-2010 genomic era confirmed mid-20th-century behavior genetics while demolishing the candidate-gene paradigm; assortative mating and genetic nurture are actively rewriting older interpretations.

## TLDR

Virtually every measured psychological trait is moderately to substantially heritable, hyper-polygenic, and shaped by environments that are themselves partly genetic in origin. The post-2010 genomic era confirmed the core findings of mid-20th-century behavior genetics while simultaneously demolishing the candidate-gene paradigm that dominated psychiatry from 1996–2010. Twin heritability for psychological traits averages ~49% (Polderman et al. 2015); molecular GWAS increasingly accounts for this through thousands of tiny-effect common variants plus rarer large-effect variants in neurodevelopmental conditions.

A crucial methodological development since 2018 is the recognition that assortative mating and gene-environment correlation systematically inflate GWAS-derived estimates. Border et al. (2022, *Science*) showed that cross-trait assortative mating alone can account for substantial fractions of reported genetic correlations — including some psychiatric cross-disorder correlations previously attributed to shared biology. Kong et al.'s (2018) "genetic nurture" finding demonstrated that roughly half of population-level polygenic score prediction for educational attainment reflects environmentally-mediated parental effects, not direct genetic causation. These corrections don't eliminate genetic influence — they reframe it.

The field's most contested findings are not the ones most disputed in public discourse: heritability is settled science, the "parenting wars" are largely resolved, and the candidate-gene-by-environment literature has collapsed. What remains genuinely open is mechanistic — how genes build minds, why the gender equality paradox exists, what drives the Flynn Effect's reversal, and whether between-population mean differences have any genetic component (a question currently unanswerable with available methods, not "settled" in either direction). The generating function for psychological variation is not "genes vs. environment" but a tightly coupled developmental system in which genetic predispositions, environments created by genetically-similar parents, assortative mating patterns, stochastic noise, and cultural context are deeply entangled.

This document is structured for someone building a formal model of psychological variation. Each section flags effect sizes, replication status, consensus, live debate, and ideological distortion from any direction.

---

## 1. Heritability: The Foundation Finding

### The Polderman meta-analysis

**Polderman et al. (2015, *Nature Genetics*)** meta-analyzed 50 years of twin research — 17,804 traits, 14.5 million twin pairs, 2,748 publications — and reported a mean heritability across all human traits of **49%**. [Polderman et al., 2015](https://www.nature.com/articles/ng.3285). For ~69% of traits, simple additive ACE models fit cleanly.

### Turkheimer's Laws and the Fourth Law

**Turkheimer's Three Laws** (2000) — all human behavioral traits are heritable; shared family environment is smaller than genes; substantial variance is explained by neither — were extended by **Chabris et al. (2015)** with the **Fourth Law: a typical behavioral trait is associated with very many genetic variants of tiny effect**. This emerged from the failure of candidate-gene studies and the polygenic architecture revealed by GWAS.

### What heritability actually means (and doesn't)

Heritability is a **population statistic**, not an individual one. Saying IQ is 70% heritable does *not* mean 70% of any person's IQ comes from genes. It is **not deterministic** (height is ~80% heritable yet rose ~10cm in 20th-century Europe through nutrition) and **not immutable** (h² changes with environment — if all environments became identical, h² would approach 1.0). The most common misinterpretation collapses statistical variance partitioning into causal mechanism.

### Twin and molecular estimates by domain

| Domain | Twin h² | SNP-h² | Largest GWAS | Loci | Best PGS R² |
|---|---|---|---|---|---|
| Adult IQ / g | 0.70–0.80 | ~0.20 | 269,867 (Savage 2018) | 205 | ~0.05 |
| Educational attainment | ~0.40 | ~0.13 | 3M (Okbay 2022) | 3,952 | 0.12–0.16 |
| Big Five (avg) | 0.40–0.60 | 0.05–0.18 | 449k (Nagel 2018) | 136 (N) | &lt;0.05 |
| Political orientation | ~0.40 | — | — | — | — |
| Religiosity | 0.30–0.45 | — | — | — | — |
| Risk tolerance | ~0.30 | 0.05 | 1M (Karlsson Linnér) | 99 | &lt;0.02 |
| Schizophrenia | 0.60–0.80 | 0.24 | 320k (Trubetskoy 2022) | 287 | 0.07–0.10 |
| Bipolar | 0.70–0.85 | 0.18–0.20 | 414k (Mullins 2021) | 64 | 0.04 |
| MDD | 0.35–0.40 | 0.09 | 807k (Howard 2019) | 102 | 0.02–0.03 |
| ADHD | 0.74 | 0.14 | 225k (Demontis 2023) | 27 | 0.04–0.06 |
| Autism | 0.80 | 0.12 | 46k (Grove 2019) | 5 | &lt;0.03 |

Note: Political orientation and religiosity are included because they are among the few adult traits where shared family environment (C) remains substantial (~20–30%), unlike personality and cognition where C ≈ 0 by adulthood. See [Alford, Funk & Hibbing (2005)](https://digitalcommons.unl.edu/cgi/viewcontent.cgi?article=1006&context=poliscifacpub); [Hatemi et al. (2014)](https://pubmed.ncbi.nlm.nih.gov/25994545/).

### The Wilson Effect

**Bouchard (2013)** documented that IQ heritability *rises* with age — from ~20% at age 5 to ~80% by adulthood, with shared-environment effects dropping from ~55% in early childhood to roughly zero by adolescence. [Bouchard 2013](https://pubmed.ncbi.nlm.nih.gov/23919982/). Briley & Tucker-Drob (2013) explained the mechanism: early genetic effects are amplified across development through gene-environment correlation (niche-picking). [Briley & Tucker-Drob 2013](https://journals.sagepub.com/doi/abs/10.1177/0956797613478618). This finding is robust, replicated, and counterintuitive — genetic differences become *more* expressed as people age into self-selected environments.

### The "missing heritability" problem

The gap between twin h² (~0.70 for IQ) and SNP-h² (~0.20) launched a decade of debate. **Wainschtein et al. (2022, *Nature Genetics*)** essentially closed it for height: using whole-genome sequencing in 25,465 unrelated individuals, h² recovered to 0.68 when rare and low-LD variants were included. [Wainschtein et al. 2022](https://www.nature.com/articles/s41588-022-01012-3). For psychological traits the same pattern is emerging. The current synthesis: missing heritability is partly real (rare variants, dominance, GxE) and partly artifactual (twin overestimation from assortative mating and rGE, measurement noise).

### Assortative mating: a pervasive inflation source

**Assortative mating (AM)** — the tendency for partners to resemble each other on traits — has emerged as a major methodological concern. People mate assortatively on education (spousal r ≈ 0.40–0.60), IQ (~0.40), personality (~0.10–0.20), height (~0.20), and psychiatric conditions. AM has three consequences for genetic estimates:

1. **Inflated heritability**: AM increases additive genetic variance across generations by creating linkage disequilibrium among causal variants. Most twin studies *underestimate* heritability by ignoring AM (counterintuitively); GWAS-based SNP-h² may be *inflated* by AM-induced LD. [Border et al. 2022, *Nat Commun*](https://pubmed.ncbi.nlm.nih.gov/35115539/).
2. **Inflated genetic correlations**: **Border et al. (2022, *Science*)** introduced **cross-trait assortative mating (xAM)** and showed that phenotypic cross-mate correlations explain R² = 74% of the variance in reported genetic correlation estimates. Some psychiatric cross-disorder genetic correlations — previously interpreted as evidence of shared biology — may be largely or entirely attributable to xAM. [Border et al. 2022](https://www.science.org/doi/10.1126/science.abo2059).
3. **Inflated PGS prediction**: Within-family PGS effects are roughly half of population-level effects for educational attainment (Okbay 2022), partly because AM and population stratification inflate between-family comparisons.

**Plomin (2022, *Behav Genet*)** argues this is a prediction-vs-explanation distinction: AM inflates causal genetic estimates but doesn't invalidate PGS as predictors, since AM-induced variance is real population variance. [Plomin 2022](https://pmc.ncbi.nlm.nih.gov/articles/PMC8960389/). This is technically correct but sidesteps the question of *why* PGS predict — whether through direct genetic causation or through correlated environments created by assortatively-mating parents.

### The genetic nurture revolution

**Kong et al. (2018, *Science*)** used 21,637 Icelandic probands with parental genotypes to compute polygenic scores from *non-transmitted* parental alleles. The non-transmitted PGS predicted offspring educational attainment at ~30% the magnitude of transmitted PGS — meaning **parental genotypes shape children via environments they create, even for alleles never inherited**. [Kong et al. 2018](https://www.bdi.ox.ac.uk/publications/823004). **Okbay et al. (2022, EA4, *Nature Genetics*)** confirmed this in 3 million people: within-family direct effects are roughly *half* the population-level PGS magnitude. [Okbay et al. 2022](https://www.nature.com/articles/s41588-022-01016-z). The implication: GWAS effect sizes for socially-valued traits are inflated by indirect/dynastic effects, and roughly half of what we used to call "genetic transmission" is actually environmentally mediated by genetically-similar parents.

---

## 2. Environmental Shaping: Real, But Smaller and Weirder Than Common Sense Suggests

The shared-vs-non-shared distinction is the single most disorienting finding for laypeople. Across hundreds of twin and adoption studies, the **shared family environment (C) accounts for ~0% of variance in adult personality and most adult cognition** ([Plomin & Daniels 1987](http://www.simine.com/240/readings/Plomin_et_al_(9).pdf); Bouchard & McGue 2003). Parental warmth, parenting style, dinner conversations, books in the home — once genetic transmission is controlled, almost none of this leaves a measurable trace on adult personality. **Important exceptions where C remains substantial: educational attainment (~20%), antisocial behavior, religiosity (~25%), political orientation (~20–30%), and childhood (but not adult) externalizing.**

### What "non-shared environment" actually is

Turkheimer & Waldron's (2000) meta-analysis of *measured* non-shared environmental predictors found these accounted for only ~2% of variance in outcomes. Plomin's recent verdict: non-shared environment is "real but largely random," more akin to **stochastic developmental noise** — differential peer experiences, illness, idiosyncratic events, measurement error — than systematic experience.

### The Equal Environments Assumption

Critics (Joseph, Charney; [Fosse et al. 2015](https://www.sciencedirect.com/science/article/abs/pii/S0049089X13001397)) note that MZ twins are treated more similarly than DZ twins. The central empirical defense: **Kendler et al. (1993)** showed misperceived-zygosity twins had phenotypic similarity tracking *true* zygosity. [Kendler et al. 1993](https://link.springer.com/article/10.1007/BF01067551). MZ-reared-apart correlations (Bouchard's Minnesota study) closely match MZ-reared-together correlations. **Felson (2014)** reanalysis: EEA is "not strictly valid, but bias is modest." Modern SNP-based heritability estimates entirely bypass EEA and give somewhat lower but still substantial h². **Verdict: EEA is approximately valid; bias is modest (~10–20% inflation) for most traits.**

### Environmental factors with robust causal effects on cognition

A small number of environmental insults have **large, replicated, causal effects** — typically asymmetric (removing severe deficits matters more than enrichment above normal):

- **Lead exposure**: Lanphear et al. (2005) pooled 7 prospective cohorts: blood lead 1→10 µg/dL → **−6.2 IQ points**, with steeper slope at low concentrations. Causal status: strong.
- **Severe iodine deficiency**: 8–12 IQ point cost; supplementation recovers ~8.7 points ([Bougma 2013](https://www.ncbi.nlm.nih.gov/pmc/articles/PMC3705354/); [Qian 2005](https://pubmed.ncbi.nlm.nih.gov/15734706/)). RCT-supported, globally replicated.
- **Heavy prenatal alcohol**: FAS produces mean IQ ~70; Mendelian-randomization confirms causation.
- **Schooling**: **Ritchie & Tucker-Drob (2018)** meta-analyzed 142 effect sizes / 600,000 participants across three quasi-experimental designs. **Each year of education raises IQ by 1–5 points (mean ~3.4)**, persisting into old age. The most consistent durable IQ-raising intervention identified. [Ritchie & Tucker-Drob 2018](https://journals.sagepub.com/doi/10.1177/0956797618774253).
- **Air pollution (PM2.5)**: ~−0.27 IQ points per 1 µg/m³ ([Aghaei 2024](https://ui.adsabs.harvard.edu/abs/2024EnvHe..23..101A/abstract)); smaller per-unit than lead but exposure is widespread.

### The Scarr-Rowe interaction (SES × heritability)

**Turkheimer et al. (2003)** reported that in impoverished families, IQ heritability was ~10% with shared environment ~60%; in affluent families this reversed. [Turkheimer et al. 2003](https://journals.sagepub.com/doi/abs/10.1046/j.0956-7976.2003.psci_1475.x). **Tucker-Drob & Bates (2016)** meta-analysis: replicated in U.S. samples but **absent in Western European/Australian samples** — likely because more universal healthcare/education reduces environmental variance at the bottom. Status: real but context-dependent and U.S.-specific.

### Parenting effects: the Harris correction, partially reversed

**Judith Rich Harris (1995; *The Nurture Assumption* 1998)** argued that within-normal-range parenting has minimal long-term effects on adult personality. The empirical core was correct: C ≈ 0 for adult personality. But Harris overstated her case. Korean-American adoption studies (**Sacerdote 2007**; [Beauchamp et al. 2023](https://papers.ssrn.com/sol3/papers.cfm?abstract_id=4491976)) show real but modest causal effects of family environment on educational attainment, BMI, drinking, smoking — transmission coefficients ~25% of biological-family magnitude. Severe deprivation/abuse causes clear damage. **The accurate position: within the normal Western range, parenting style has small effects on adult personality; family environment has measurable but modest effects on attainment outcomes; severe parenting variation matters substantially.**

### Neighborhood and peer effects

**Chetty & Hendren (2018, *QJE*)** used 5+ million U.S. cross-county movers with sibling fixed effects: each year of childhood exposure to a 1-SD better county raises adult income by ~0.5–0.7%. [Chetty & Hendren 2018](https://opportunityinsights.org/paper/neighborhoodsi/). **Moving to Opportunity reanalysis** ([Chetty, Hendren & Katz 2016](https://www.nber.org/papers/w21156)): children moving before age 13 had adult earnings 31% higher than controls. Place matters, but accumulates slowly across many years of exposure.

### The Flynn Effect and its reversal

Flynn (1984, 1987) documented ~3 IQ points/decade gains across the 20th century. Causes contested: nutrition, schooling, infectious-disease reduction, test sophistication, smaller families — no single mechanism established. **Bratsberg & Rogeberg (2018, *PNAS*)** used Norwegian within-family conscript data to demonstrate that both the Flynn Effect and its **post-1990s reversal** are environmentally driven (visible within sibships, ruling out dysgenic/compositional explanations). [Bratsberg & Rogeberg 2018](https://www.pnas.org/doi/10.1073/pnas.1718793115). Similar declines now reported in Denmark, Finland, the Netherlands, France, the UK, and Germany. The reversal's cause is unknown — this is one of the field's most important open questions.

---

## 3. Gene-Environment Interplay: rGE Wins, Candidate-GxE Collapsed

### Three types of gene-environment correlation

The Plomin/DeFries/Loehlin (1977) framework distinguishes **passive rGE** (parents transmit both genes and correlated rearing environment), **evocative rGE** (heritable child traits elicit specific responses), and **active rGE / niche-picking** (individuals select environments matching genetic propensities). **Kendler & Baker's (2007)** systematic review shows essentially every measured environment is itself heritable (15–35%) — meaning observational claims like "parental warmth causes child outcomes" are confounded by passive rGE. The genetic nurture and within-family PGS results (Section 1) quantify this: **population-level "genetic" prediction is roughly half indirect environmental effects of genetically-similar parents**.

### The candidate gene × environment collapse

**Caspi et al. (2003, *Science*)** reported that 5-HTTLPR short-allele carriers showed elevated depression risk under stress. The paper became one of the most cited in psychiatry (>9000 citations). It collapsed:

- **Risch et al. (2009, *JAMA*)**: meta-analysis of 14 studies, N=14,250 — no evidence. [Risch et al. 2009](https://www.ncbi.nlm.nih.gov/books/NBK76610/).
- **Culverhouse et al. (2018)**: pre-registered collaborative meta-analysis, 31 datasets, N=38,802 — definitively no evidence.
- **Border et al. (2019, *Am J Psychiatry*)**: examined 18 most-studied depression candidate genes in N up to 443,264. **No clear evidence** for any candidate gene polymorphism on depression. As a set, candidate genes were no more associated with depression than non-candidate genes. [Border et al. 2019](https://profiles.wustl.edu/en/publications/no-support-for-historical-candidate-gene-or-candidate-gene-by-int/).
- **Duncan & Keller (2011)**: 96% of novel candidate-GxE studies were significant; only 27% of replication attempts were. [Duncan & Keller 2011](https://pmc.ncbi.nlm.nih.gov/articles/PMC3222234/).

**MAOA × maltreatment** (Caspi et al. 2002) is the partial exception that survived meta-analysis ([Byrd & Manuck 2014](https://pmc.ncbi.nlm.nih.gov/articles/PMC3858396/)) — modest male-specific interaction, but smaller than originally reported.

### Differential susceptibility / orchid-dandelion

**Belsky & Pluess (2009)** reframed "risk alleles" as "plasticity alleles" — some individuals are more reactive to environments "for better and for worse." [Belsky & Pluess 2009](https://pubmed.ncbi.nlm.nih.gov/19883141/). The theory is generative; the empirical record is mixed. Recent systematic reviews find that interactions between child characteristics and parenting rarely replicate across cohorts and developmental domains. Distinguishing differential susceptibility from diathesis-stress requires very large, preregistered samples. [de Villiers et al. 2018](https://pubmed.ncbi.nlm.nih.gov/30626458/).

### Epigenetics: real biology, oversold psychology

DNA methylation, histone modifications, and non-coding RNA regulation are real, well-characterized mechanisms important in development. The controversy concerns whether environmentally-induced epigenetic marks are faithfully transmitted across generations in humans. **They generally are not.**

- **Heard & Martienssen (2014, *Cell*)**: in mammals, two waves of near-complete epigenetic reprogramming erase most acquired methylation marks. Robust transgenerational epigenetic inheritance occurs in plants and *C. elegans*; in humans it remains largely speculative. [Heard & Martienssen 2014](https://www.sciencedirect.com/science/article/pii/S0092867414002864).
- **Dutch Hunger Winter** ([Heijmans et al. 2008](https://www.pnas.org/doi/10.1073/pnas.0806560105)): real *within-individual* epigenetic effect persisting decades, *not* evidence of transmission to grandchildren.
- **Yehuda's Holocaust FKBP5 study** (2016): tiny sample (n=8 control parents), opposite-direction effects in parents vs. offspring, no germline measurement. Yehuda's own group failed to replicate. The "trauma is inherited epigenetically" narrative is not supported by current evidence.

### Critical periods: solid developmental neuroscience

**Hensch (2005, *Nat Rev Neurosci*)** provides a mechanistically rigorous account of cortical critical-period plasticity. [Hensch 2005](https://www.nature.com/articles/nrn1787). GABAergic maturation (parvalbumin-positive interneurons) gates onset; perineuronal nets and myelin-associated inhibitors close periods. This represents the high end of how environmental experience shapes brain structure — genuine, replicated, and mechanistically understood.

---

## 4. Sex and Gender Differences: Large Where You're Not Told They Are

Sex differences are one of psychology's most ideologically distorted areas — distorted by both minimization and overstatement. The actual picture: **small differences in average cognitive ability, large differences in interests and physical aggression, moderate-to-large multivariate personality differences, and a robust but mechanistically contested gender equality paradox.**

### Cognitive abilities

Mental rotation shows **d ≈ 0.56–0.73 male advantage** (Voyer et al. 1995), among the largest cognitive sex differences documented. Mean math performance: **d ≈ 0.05–0.10** ([Lindberg et al. 2010](https://pmc.ncbi.nlm.nih.gov/articles/PMC3057475/)) — essentially no average difference. Writing: substantial female advantage. School grades favor girls overall ([Voyer & Voyer 2014](https://pubmed.ncbi.nlm.nih.gov/24773502/)). At extreme tails (95th–99th percentile) males outnumber females ~2:1 in many countries — driven by **slightly greater male variance** (~3–15% higher) compounding at extremes.

### Personality: univariate vs. multivariate framing

Univariate Big Five differences are moderate: women higher on Neuroticism (d ≈ 0.40) and Agreeableness (d ≈ 0.40). **Del Giudice, Booth & Irwing (2012)** computed multivariate **Mahalanobis D = 2.71** on 16PF data from 10,261 Americans, implying ~10% overlap between male and female personality profiles. [Del Giudice et al. 2012](https://journals.plos.org/plosone/article/file?id=10.1371/journal.pone.0029265&type=printable). Hyde's (2005) "Gender Similarities Hypothesis" — most differences trivial or small — is mathematically compatible but tells a very different qualitative story. **Both univariate and multivariate framings should be reported jointly**; selective use is ideological.

### Interests: the largest sex difference in psychology

**Su, Rounds & Armstrong (2009, *Psych Bulletin*)** meta-analyzed 503,188 people: the **People-Things dimension d = 0.93**, with engineering interest d = 1.11. [Su et al. 2009](https://eric.ed.gov/?id=EJ860932). These are *very large* by psychological standards and the largest in the entire literature on psychological sex differences.

### Aggression

Archer (2004): physical aggression d ≈ 0.40–0.60 male; trait anger near zero. Males commit ~95% of homicides globally. [Archer 2004](https://journals.sagepub.com/doi/10.1037/1089-2680.8.4.291). Indirect/relational aggression: Card et al. (2008) found differences trivial (d &lt; 0.10), challenging the "girls do indirect aggression equally" narrative.

### The Gender Equality Paradox (replicated; mechanism contested)

A robust empirical pattern across at least four domains: **personality, preference, interest, and depression-rate differences are larger in more gender-equal and wealthier countries.**

- **Schmitt et al. (2008)**: 55-nation Big Five study — differences largest in egalitarian Western cultures. [Schmitt et al. 2008](https://pubmed.ncbi.nlm.nih.gov/18712468/).
- **Falk & Hermle (2018, *Science*)**: 80,000 adults, 76 countries — sex differences in 6 economic preferences positively related to GDP and gender equality.
- **Stoet & Geary (2018)**: STEM Gender-Equality Paradox — more gender-equal countries had *smaller* female share of STEM graduates. A corrigendum addressed methods; the core correlation remained robust.

The correlation is robust. The causal mechanism — innate-expression release in wealthy environments vs. measurement artifacts vs. ecological confounds — is genuinely contested.

### Mental health asymmetries

Depression female:male ≈ 2:1; anorexia ~10:1 female; ADHD diagnosis ~2–3:1 male; antisocial personality, substance use, completed suicide all male-skewed; autism ~3–4:1 male; schizophrenia roughly equal but more severe early-onset in males.

### Biological mechanisms

CAH girls (prenatally elevated androgens) show masculinized toy preferences and play patterns ([Kung et al. 2024 meta-analysis](https://www.sciencedirect.com/science/article/abs/pii/S014976342400085X)). Same-sex-typed toy preferences in vervet and rhesus monkeys parallel human findings, supporting partial biological mediation. Wood & Eagly's social role theory faces empirical challenge from the gender equality paradox.

---

## 5. Cognitive Ability and Intelligence

### The g-factor

Spearman's 1904 finding of a **positive manifold** — every cognitive test correlates positively with every other — is arguably the most replicated finding in psychology. A first unrotated principal factor captures 40–50% of variance in any sufficiently broad battery. **van der Maas et al. (2006)** mutualism model offers an alternative: g may be an emergent network property of reciprocally beneficial cognitive processes during development, not a unitary biological cause. [van der Maas et al. 2006](https://pubmed.ncbi.nlm.nih.gov/17014305/). Most working researchers treat g as a robust statistical regularity whose causal architecture is unsettled.

### Structure: CHC theory

**Carroll's (1993)** three-stratum theory — g at top, ~8–10 broad abilities (Gf, Gc, Gv, Ga, Gs, Gsm, Glr, Gq, Grw), ~70+ narrow abilities — was integrated with Cattell-Horn into the **Cattell-Horn-Carroll (CHC)** framework, which underlies modern IQ tests.

### Predictive validity

**Schmidt & Hunter (1998)**: corrected GMA validity for job performance r ≈ 0.51. **Sackett et al. (2022)** argued corrections were too aggressive; re-estimate: **r ≈ 0.31 uncorrected / ~0.42 corrected**. GMA remains among the most predictive selection tools. Childhood IQ predicts educational attainment at r ≈ 0.50–0.70. **Calvin et al. (2011)** meta-analysis (1.1M, 22,453 deaths): each 1-SD higher childhood IQ → ~24% lower all-cause mortality. [Calvin et al. 2011](https://pubmed.ncbi.nlm.nih.gov/21037248/).

### Lifespan stability

**Lothian Birth Cohort**: age 11 → age 90 corrected correlation r ≈ 0.67. [Deary et al. 2013](https://pubmed.ncbi.nlm.nih.gov/24084038/). About one-third of variance in mental ability at 90 is accounted for by ability at 11.

### Group differences in test scores: the most distorted area

**Roth et al. (2001)** meta-analysis (N=6.2M): U.S. Black-White cognitive ability gap **d ≈ 1.0 (~15 IQ points)**. **Dickens & Flynn (2006)**: Black IQ rose 4–7 points relative to whites between 1972–2002 (about one-third of the gap). [Dickens & Flynn 2006](https://journals.sagepub.com/doi/abs/10.1111/j.1467-9280.2006.01802.x). The gap exists, has narrowed somewhat, and has not closed.

The mainstream contemporary position (**Nisbett et al. 2012**; **Turkheimer, Harden & Nisbett 2017**): within-group heritability does *not* license between-group inferences (Lewontin's point); **Martin et al. (2019, *Nature Genetics*)** demonstrated PGS lose ~4.5x prediction accuracy in African-ancestry individuals due to differential LD and allele frequencies, meaning **current PGS cannot validly compare mean genetic predisposition across continental ancestry groups**. [Mostafavi et al. (2020)](https://elifesciences.org/articles/48376) showed PGS portability also breaks down *within* Europeans across SES strata.

**The honest scientific position: gaps in test scores are real, partly narrowing, and their causes are not currently identifiable as genetic, environmental, or both** — direct evidence is absent and mainstream geneticists treat the question as not currently answerable.

**Distortion from the hereditarian direction**: treating g-loadedness as evidence of genetic etiology (environmental causes can also be g-loaded); citing fringe admixture studies published in weak-peer-review venues; conflating absence of evidence with agnosticism. **Distortion from the environmentalist direction**: claiming gaps have closed when they only partly narrowed; dismissing IQ as "culturally biased" despite measurement-invariance evidence; overstating stereotype threat (Flore & Wicherts 2015 meta-analysis showed publication-biased modest effects).

### Brain correlates

Brain volume × IQ: r ≈ 0.24 ([Pietschnig et al. 2015, 2022](https://royalsocietypublishing.org/rsos/article/9/5/211621/96643/Of-differing-methods-disputed-estimates-and)). **P-FIT theory** (Jung & Haier 2007): intelligence supported by parieto-frontal network. [Jung & Haier 2007](https://www.cambridge.org/core/journals/behavioral-and-brain-sciences/article/parietofrontal-integration-theory-pfit-of-intelligence-converging-neuroimaging-evidence/175ECF09153B51136C21325465C6A86F).

### Creativity and intelligence

The "IQ ≈ 120 threshold" hypothesis is largely disconfirmed ([Weiss et al. 2020](https://pmc.ncbi.nlm.nih.gov/articles/PMC7709632/)). Intelligence and creativity correlate ~r = 0.20–0.30 across the range. Openness to Experience is the personality trait most reliably correlated with creative achievement (~0.30–0.40).

---

## 6. Personality and Temperament

### The Big Five (OCEAN) and HEXACO

The Big Five emerged from the lexical hypothesis. **Heritability is ~40–60% per twin studies**; SNP-h² is 8–18%. **Nagel et al. (2018)** identified 136 loci for neuroticism in 449,484 people. [Nagel et al. 2018](https://pubmed.ncbi.nlm.nih.gov/29942085/). The ReGPC consortium (2025) reports 703 loci for neuroticism in 1M+ participants. [ReGPC 2025](https://www.biorxiv.org/content/10.1101/2025.05.16.648988.full.pdf).

**Roberts & DelVecchio (2000)**: rank-order stability rises from ~0.31 in childhood to ~0.74 by midlife (**cumulative continuity**). [Roberts & DelVecchio 2000](http://jenni.uchicago.edu/Spencer_Conference/Representative%20Papers/Roberts%20&%20DelVecchio,%202000.pdf). **Roberts et al. (2006)** the **maturity principle**: mean-level increases in Conscientiousness, Agreeableness, and Emotional Stability with age, especially in young adulthood. [Bleidorn et al. 2022 update](https://www.researchgate.net/publication/362021332_Personality_stability_and_change_A_meta-analysis_of_longitudinal_studies).

**HEXACO** (Ashton & Lee): lexical studies in 12+ languages consistently yield six factors, the sixth being **Honesty-Humility**. H predicts integrity-related criteria incrementally over Big Five. [Ashton & Lee 2008](https://compass.onlinelibrary.wiley.com/doi/abs/10.1111/j.1751-9004.2008.00134.x).

### Temperament: the developmental foundation

Temperament research constitutes a parallel tradition to adult personality, focused on biologically-grounded individual differences emerging in infancy.

**Rothbart's model** identifies three overarching dimensions: **Surgency/Extraversion** (activity, positive affect, approach), **Negative Affectivity** (fear, anger, sadness, discomfort), and **Effortful Control** (attentional regulation, inhibitory control, low-intensity pleasure). Effortful Control is particularly important — it is the self-regulatory component of temperament, developing primarily during ages 2–7 as the anterior attention network matures, and is a strong predictor of later externalizing problems, academic success, and conscience development.

**Kagan's Behavioral Inhibition (BI)** framework focuses on extreme phenotypes: ~15–20% of infants show high-reactive patterns (vigorous motor activity and distress to novel stimuli at 4 months) who become behaviorally inhibited toddlers — cautious, avoidant with unfamiliar people and situations. BI maps approximately onto low Surgency + high Negative Affectivity (especially fear). Kagan's longitudinal studies showed BI is moderately heritable (~50%), associated with higher resting heart rate and amygdala excitability, and predicts elevated risk for social anxiety disorder in adolescence (OR ~2–4). However, ~60% of high-reactive infants do *not* become clinically anxious adults — biology is a foundation, not a constraint.

**Thomas & Chess's (1977)** "goodness of fit" model — later empirically supported — emphasized that temperamental difficulty per se doesn't predict poor outcomes; the match between child temperament and environmental demands does.

The temperament → personality continuity is increasingly well-documented: infant Surgency maps onto adult Extraversion; infant Negative Affectivity onto Neuroticism; infant Effortful Control onto Conscientiousness. The mapping is imperfect — adult personality includes social-cognitive layers (identity, values, narrative) absent in temperament.

### Cross-cultural universality

**McCrae & Terracciano (2005)**: clean Big Five replication in 50 cultures. [McCrae & Terracciano 2005](https://www.researchgate.net/publication/7994365_Universal_Features_of_Personality_Traits_From_the_Observer's_Perspective_Data_From_50_Cultures). **Gurven et al. (2013)** challenged this with the Tsimane forager-horticulturalists, where the full Big Five did not robustly emerge. [Gurven et al. 2013](https://pubmed.ncbi.nlm.nih.gov/23245291/). **Consensus**: 3 factors (E, A, C) replicate cross-linguistically; the full Big Five replicates well in Indo-European languages; non-WEIRD samples sometimes show structural deviations.

### Dark traits and the D factor

**Paulhus & Williams (2002)**: the Dark Triad (Machiavellianism, narcissism, psychopathy). [Buckels, Jones & Paulhus (2013)](https://pubmed.ncbi.nlm.nih.gov/24022650/) added everyday sadism. **Moshagen, Hilbig & Zettler (2018)** proposed the **D factor** — a general tendency to maximize individual utility while disregarding others — as the common core, mapping strongly onto low Honesty-Humility. [Moshagen et al. 2018](https://pubmed.ncbi.nlm.nih.gov/29999338/).

### Person-situation debate: resolved

The Mischel (1968) critique — cross-situational consistency rarely exceeds r ≈ 0.30 — was resolved through aggregation, interactionism (Mischel & Shoda's CAPS model), and Fleeson's within-person variability framework. The modern consensus: persons, situations, and their interactions all matter.

### Personality predicts outcomes as strongly as IQ and SES

**Roberts et al. (2007)** "The Power of Personality": meta-analytic comparison shows personality effects on mortality, divorce, and occupational attainment are *indistinguishable in magnitude* from SES and cognitive ability effects. [Roberts et al. 2007](https://journals.sagepub.com/doi/10.1111/j.1745-6916.2007.00047.x). Conscientiousness predicts mortality through health behaviors with large effect size. [Bogg & Roberts 2004](https://gwern.net/doc/psychology/personality/conscientiousness/2004-bogg.pdf).

### Recent theoretical developments

**DeYoung's Cybernetic Big Five Theory (2015)**: traits as parameters of a cybernetic goal-pursuit system. [DeYoung 2015](https://www.semanticscholar.org/paper/Cybernetic-Big-Five-Theory.-DeYoung/432f424abfc5891b009f89eeb3ed112e664af50e). **Mõttus et al. (2017)**: "personality nuances" research argues item-level traits capture incremental valid variance below the facet level. [Mõttus et al. 2017](https://pmc.ncbi.nlm.nih.gov/articles/PMC6348042/).

---

## 7. Neurodiversity and Psychopathology: Dimensional, Polygenic, Transdiagnostic

The genomic era has produced three conclusions that fundamentally reshape psychiatric nosology: **all major psychiatric conditions are highly heritable, hyper-polygenic, and substantially genetically overlapping across diagnostic categories.**

### Headline findings by disorder

- **Schizophrenia**: twin h² ~80%; **Trubetskoy et al. (2022, *Nature*)**: **287 loci**; SNP-h² ~24%. [Trubetskoy et al. 2022](https://www.nature.com/articles/s41586-022-04434-5). Environmental risk factors: urban birth (~2× risk), high-potency cannabis (OR ~3.9), migration, obstetric complications.
- **Bipolar**: twin h² ~70–85%; **Mullins et al. (2021)**: **64 loci**; rg(SCZ,BD) ~0.7. [Mullins et al. 2021](https://pmc.ncbi.nlm.nih.gov/articles/PMC8192451/).
- **Major Depression**: h² ~37%; **Howard et al. (2019)**: **102 loci**; SNP-h² ~9%. [Howard et al. 2019](https://pmc.ncbi.nlm.nih.gov/articles/PMC6522363/). Strong rg with neuroticism (~0.7).
- **ADHD**: twin h² ~74%; **Demontis et al. (2023)**: **27 loci**. [Demontis et al. 2023](https://www.researchgate.net/publication/358716011_Genome-wide_analyses_of_ADHD_identify_27_risk_loci_refine_the_genetic_architecture_and_implicate_several_cognitive_domains). Negative rg with educational attainment and IQ.
- **Autism**: twin h² ~80%; **Grove et al. (2019)**: **5 common-variant loci** plus substantial rare/de novo variants of large effect (CHD8, SCN2A, SYNGAP1). [Grove et al. 2019](https://www.nature.com/articles/s41588-019-0344-8). Common-variant PGS positively correlated with IQ and education; ID-comorbid autism (rare-variant-driven) negatively correlated.

### Cross-disorder pleiotropy (with assortative mating caveat)

**Brainstorm Consortium (2018, *Science*)**: substantial genetic correlations among psychiatric disorders. **Cross-Disorder PGC (Lee et al. 2019, *Cell*)**: across 8 disorders, **109 pleiotropic loci**, three clusters — compulsive, mood/psychotic, early-onset neurodevelopmental. [Lee et al. 2019](https://www.biorxiv.org/content/10.1101/528117v1).

**Critical caveat**: Border et al. (2022, *Science*) showed that cross-trait assortative mating can generate spurious genetic correlations between phenotypes with entirely distinct genetic bases. Some fraction of reported psychiatric cross-disorder genetic correlations may reflect xAM rather than shared biology. The magnitude of this artifact is actively being quantified and represents a major revision in progress.

### The p-factor

**Caspi et al. (2014)**: a single **p (general psychopathology)** factor fit Dunedin cohort data better than three-factor models — analogous to g for cognitive ability. [Caspi et al. 2014](https://journals.sagepub.com/doi/abs/10.1177/2167702613497473). Higher p associated with greater impairment, familiality, worse developmental histories. Replicated in dozens of samples. Interpretations contested: genuine common liability, statistical artifact of bifactor over-extraction, or a reflection of impairment/distress per se.

### Dimensional alternatives: HiTOP and RDoC

**HiTOP** ([Kotov et al. 2017](https://pubmed.ncbi.nlm.nih.gov/33577350/)): a quantitatively-derived dimensional alternative to DSM organized hierarchically. **RDoC** (NIMH 2009–): six dimensional neurobiologically-grounded research domains. Both converge with taxometric evidence (most psychopathology is dimensional, not taxonic) on the **dimensional turn in psychiatric science**.

### Polygenic scores in clinics: not yet

Best PGS R² ~7–10% for schizophrenia. PGS alone does not outperform family history. PGS performance drops 50–70% in non-European-ancestry populations — a major equity and portability problem.

### The neurodiversity framework: scientific–identity tensions

Coined by Singer (1998), the neurodiversity paradigm reframes autism, ADHD, dyslexia as natural variation rather than pathology. The framework has legitimate ethical force but operates in tension with deficit-oriented findings for severe presentations (profound autism with ID, epilepsy, self-injury). A defensible position recognizes both the reality of impairment at the severe end and the population-level continuous variation that grades into normality.

---

## 8. Key Researchers and Labs

| Researcher | Affiliation | Central contribution |
|---|---|---|
| **Robert Plomin** | King's College London | Behavioral genetics synthesis; Blueprint; GPS |
| **Eric Turkheimer** | University of Virginia | Three Laws; Scarr-Rowe; philosophical foundations |
| **K. Paige Harden** | UT Austin | Genetic Lottery; causal inference with PGS |
| **Avshalom Caspi / Terrie Moffitt** | Duke / King's | p-factor; Dunedin cohort; (and candidate-GxE) |
| **Ian Deary** | Edinburgh | Lothian Birth Cohorts; cognitive epidemiology |
| **Elliot Tucker-Drob** | UT Austin | Education-IQ meta-analysis; Wilson Effect mechanisms |
| **Daniel Benjamin / SSGAC** | UCLA | EA GWAS consortium; social-science genomics |
| **Colin DeYoung** | Minnesota | Cybernetic Big Five Theory; personality neuroscience |
| **Jay Belsky** | UC Davis | Differential susceptibility |
| **Marco Del Giudice** | UNM | Multivariate sex differences |
| **Janet Hyde** | Wisconsin | Gender similarities hypothesis |
| **David Geary** | Missouri | Sex differences in math/STEM |
| **Brent Roberts** | UIUC | Personality development; maturity principle |
| **Alexander Young** | UCLA | Genetic nurture; within-family methods |
| **Richard Border** | Harvard/UCLA | Candidate gene demolition; xAM |
| **Peter Hatemi / John Hibbing** | Penn State / Nebraska | Genopolitics; heritability of political attitudes |

---

## 9. The Integrated Picture: What Generates Psychological Variation

### The model

A formal model of individual psychological variation should treat the person as the joint product of:

**(a) A hyper-polygenic genome** encoding thousands of small-effect predispositions (plus some rare large-effect variants in neurodevelopmental conditions). Twin h² for most traits falls in 0.40–0.80.

**(b) Substantial gene-environment correlation** through passive (parents transmit genes + correlated environments), evocative (child traits elicit responses), and active (niche-picking) channels. Roughly half of population-level PGS prediction reflects indirect/environmental mediation by genetically-similar parents, not direct genetic causation.

**(c) Assortative mating** inflating additive genetic variance, genetic correlations between traits, and PGS prediction accuracy. This is a recently-quantified source of systematic bias in nearly all genetic estimates.

**(d) A small set of large-effect environmental insults** — lead, severe iodine deficiency, heavy prenatal alcohol, severe deprivation — plus schooling (~3.4 IQ points/year). Effects are typically asymmetric: removing severe deficits matters more than enriching above-normal environments.

**(e) Substantial stochastic developmental noise** — the dominant source of the non-shared environment, which accounts for ~50% of personality variance and is not yet well-characterized mechanistically.

**(f) Cultural/institutional contexts** that modulate which genetic predispositions are expressed and rewarded (WEIRD effects, gender equality paradox, Scarr-Rowe interaction, Flynn Effect).

**(g) Developmental unfolding across time** — temperament in infancy (biologically grounded reactivity and regulation) becomes personality in adulthood (adding social-cognitive layers), with heritability increasing across the lifespan (Wilson Effect) and rank-order stability rising to ~0.74 by midlife.

### Where political distortion is strongest, by direction

**From the environmentalist/blank-slate direction**: dismissing twin study validity wholesale; overstating Scarr-Rowe; promoting transgenerational epigenetic narratives that exceed evidence; dismissing IQ as culturally biased despite measurement-invariance findings; overstating stereotype-threat magnitudes; minimizing the gender equality paradox.

**From the hereditarian direction**: citing within-population heritability to license between-population genetic inferences; citing fringe admixture studies as if mainstream; treating g-loadedness of gaps as evidence of genetic etiology when environmental causes can also be g-loaded; ignoring the assortative mating and genetic nurture corrections to PGS.

**From the "gender similarities" direction**: selective citation of d ≈ 0.05 for math to imply no differences anywhere; obscuring multivariate D ≈ 2.71 with univariate framing; minimizing d ≈ 0.93 people-things interest differences.

**From popular evolutionary psychology**: treating dimensional differences as taxonic; extrapolating from small ds to categorical claims; overgeneralizing from specific tasks to broad domain claims.

### Open questions worth modeling

1. **Mechanistic interpretation of PGS**: Plomin's "causal genetic" view vs. Turkheimer's "weak genetic explanation" — genuinely open.
2. **Flynn Effect reversal**: cause unknown; one of the most important open questions in differential psychology.
3. **Gender equality paradox mechanism**: innate-expression release vs. measurement artifacts vs. wealth confounds — unsettled.
4. **Between-population cognitive differences**: currently scientifically unanswerable (PGS portability too poor; cross-ancestry GWAS at scale don't exist). Honest position: unresolved, not settled in either direction.
5. **The causal architecture of g**: latent common cause vs. emergent network property (mutualism) — the positive manifold is not in dispute; what generates it is.
6. **What non-shared environment actually is**: stochastic noise, epigenetic variation, immune/microbial variation, differential peer networks — largely uncharacterized despite accounting for ~50% of personality variance.
7. **Assortative mating correction magnitudes**: how much do AM and xAM corrections change the substantive picture of genetic architecture and cross-trait pleiotropy? Active area of revision.

---

## 10. Load-Bearing Assumptions and Falsification Conditions

This section makes explicit which conclusions in this review depend on which assumptions, and what evidence would substantially revise or flip them. Ordered roughly by how much of the document's picture collapses if the assumption fails.

### Assumption 1: The twin method provides approximately valid variance decomposition

**What depends on it**: Nearly all h² estimates in Section 1's table, the C ≈ 0 finding for adult personality, the Wilson Effect, the Scarr-Rowe interaction.

**Status**: Approximately valid. SNP-h² estimates (which bypass EEA entirely) give lower but still substantial heritability for every trait measured. MZ-reared-apart designs converge with MZ-reared-together. Felson (2014) estimates ~10–20% EEA-induced inflation, not enough to eliminate the core finding.

**What would flip it**: SNP-h² for psychological traits systematically converging on &lt;0.05 (would suggest twin h² is mostly EEA artifact). Or: a large, well-powered MZ-reared-apart study finding IQ correlations &lt;0.40 (current estimates ~0.70). Neither has occurred.

**Robustness verdict**: HIGH. The convergence of twin, adoption, and molecular methods on moderate-to-substantial heritability is the most replicated finding in the field.

### Assumption 2: GWAS identifies real genetic signal (not just population structure and AM artifacts)

**What depends on it**: The entire PGS enterprise, genetic nurture estimates, cross-disorder pleiotropy findings, the "missing heritability" narrative.

**Status**: Substantially valid but with known inflation. Within-family PGS effects are non-zero for educational attainment (~half of population effects), meaning direct genetic signal exists. But the magnitude of AM and stratification inflation is still being quantified.

**What would flip it**: Within-family PGS effects for most traits converging on ~zero (would mean population-level PGS prediction is entirely indirect/environmental). Current evidence: within-family effects are reduced but clearly non-zero for EA, BMI, height; less well-characterized for personality and psychiatric traits.

**Robustness verdict**: MODERATE-HIGH for the existence of direct genetic effects; MODERATE for their precise magnitude, which is actively being revised downward.

### Assumption 3: g is a real dimension of individual variation (not a measurement artifact)

**What depends on it**: The entire intelligence section (Section 5), predictive validity claims, group-difference discussions, the CHC structure.

**Status**: The positive manifold is among the most replicated findings in psychology. Whether g is a latent common cause or an emergent network property (mutualism) is unsettled, but both interpretations preserve g's predictive validity and the meaningfulness of individual differences in general cognitive ability.

**What would flip it**: A sufficiently broad, well-constructed cognitive battery where the first principal component explains &lt;15% of variance (would undermine the positive manifold). Or: successful interventions that consistently raise one cognitive ability while lowering others (would violate the manifold's structure). Neither has been demonstrated.

**Robustness verdict**: HIGH for g as a statistical regularity with predictive validity. MODERATE for g as a unitary biological mechanism (mutualism remains a viable alternative).

### Assumption 4: Sex-difference effect sizes from meta-analyses are not primarily measurement artifacts

**What depends on it**: The gender equality paradox, the claim that interest differences (d = 0.93) are among psychology's largest, the multivariate personality finding (D = 2.71).

**Status**: Interest measures (Su et al. 2009) use well-validated instruments; the d = 0.93 holds across inventories and cultures. The Del Giudice multivariate D is sensitive to the number of variables included and the specific battery, though the qualitative finding (large multivariate difference despite moderate univariate ds) is robust across datasets. CAH and non-human primate evidence provides independent convergent support for biological mediation of interest differences.

**What would flip it**: A large cross-cultural study using behavioral (not self-report) interest measures finding d &lt; 0.30 for people-things. Or: evidence that the gender equality paradox disappears when using non-self-report personality measures (reference-group effects could inflate self-report differences in egalitarian countries). Current evidence: Falk & Hermle (2018) used incentivized behavioral measures for some preferences and found the paradox held, but full behavioral replication across all domains is incomplete.

**Robustness verdict**: HIGH for the existence of substantial sex differences in interests and aggression. MODERATE for the precise magnitude of multivariate personality differences. MODERATE for the gender equality paradox's causal interpretation.

### Assumption 5: The candidate-GxE collapse generalizes — specific gene × environment interactions are mostly small or nonexistent

**What depends on it**: Section 3's dismissal of 5-HTTLPR and similar findings, the shift toward polygenic approaches.

**Status**: For candidate genes, the collapse is definitive (Border et al. 2019). But this does not necessarily mean polygenic-score × environment interactions are also null. PGS × environment work is younger, uses better methods, and could in principle yield robust results.

**What would flip it**: Multiple large, pre-registered PGS × measured-environment studies showing robust, replicable interactions explaining >5% of variance. Current evidence: a few suggestive findings (PGS-for-education × compulsory schooling reforms) but nothing approaching the scale or replication needed for confidence.

**Robustness verdict**: HIGH for the candidate-gene collapse. LOW-MODERATE confidence in the broader claim that specific GxE interactions are generally small — this is an extrapolation from the candidate-gene failure, and the polygenic GxE literature is too young to draw strong conclusions.

### Assumption 6: Cross-disorder genetic correlations reflect shared biology (pleiotropy)

**What depends on it**: The p-factor interpretation, HiTOP structure, the "dimensional turn" in psychiatry, transdiagnostic treatment rationales.

**Status**: Substantially challenged by Border et al. (2022). Cross-trait assortative mating can generate spurious genetic correlations between traits with entirely distinct genetic bases. The R² = 74% finding means most of the variance in genetic correlation *estimates* tracks spousal phenotypic correlations — though this does not prove all genetic correlations are spurious (some genuine pleiotropy surely exists).

**What would flip it**: Within-family designs showing that cross-disorder genetic correlations survive AM correction at >50% of current estimates. Or: identification of specific shared biological pathways (e.g., synaptic pruning variants affecting both SCZ and BD) that don't depend on LD induced by AM.

**Robustness verdict**: MODERATE. The dimensional/transdiagnostic pattern is likely real but inflated. The magnitude of genuine pleiotropy vs. AM artifact is one of the field's most active methodological debates.

---

## 11. Toward Topology: Structure for the Next Phase

This section identifies the natural graph/network structure embedded in this literature, to facilitate the transition from landscape analysis to formal topology mapping.

### Natural node types

1. **Trait nodes**: Cognitive abilities (g, Gf, Gc, Gv, Gs...), personality dimensions (Big Five/HEXACO factors and facets), temperament dimensions (Surgency, Negative Affectivity, Effortful Control), psychopathology spectra (internalizing, externalizing, thought disorder), interests (people-things, RIASEC), political/moral attitudes
2. **Mechanism nodes**: Genetic architecture (common polygenic, rare large-effect, de novo), environmental factors (lead, iodine, schooling, deprivation, neighborhoods), developmental processes (critical periods, niche-picking, genetic nurture, AM), stochastic noise
3. **Method nodes**: Twin studies, adoption studies, GWAS, PGS, within-family designs, Mendelian randomization, meta-analysis
4. **Population-level modifier nodes**: SES (Scarr-Rowe), culture (WEIRD), gender equality index, historical period (Flynn Effect)

### Natural edge types

1. **Genetic correlations** (with AM caveat): e.g., rg(SCZ, BD) ≈ 0.7; rg(EA, IQ) ≈ 0.7; rg(neuroticism, MDD) ≈ 0.7
2. **Developmental continuity**: temperament → personality (Surgency → Extraversion; Effortful Control → Conscientiousness)
3. **Causal environmental effects**: lead → IQ (−6.2 pts per 10 µg/dL); schooling → IQ (+3.4 pts/year)
4. **Predictive validity edges**: g → job performance (r ≈ 0.42); Conscientiousness → mortality; EA PGS → income
5. **Methodological dependency**: twin h² → SNP-h² → PGS R² (each constraining the next)
6. **Taxonomic hierarchy**: g → broad abilities → narrow abilities (CHC); p → spectra → subfactors → syndromes (HiTOP)
7. **Moderation edges**: SES × heritability (Scarr-Rowe); gender equality × sex differences (GEP); age × heritability (Wilson Effect)

### Key structural features for the graph

- **Two parallel hierarchies** (CHC for cognition, HiTOP for psychopathology) that share genetic correlations at the top level (g correlates with p inversely)
- **A developmental cascade** from temperament (infancy) through personality (adulthood) through outcomes (mortality, income, relationships), with heritability increasing and shared-environment decreasing across the lifespan
- **A methodological funnel** from twin estimates (broadest, highest h²) through molecular estimates (narrower, lower h²) through within-family estimates (narrowest, lowest but most causally clean)
- **Cross-domain genetic correlations** that form a web connecting cognition, personality, and psychopathology — but with the critical caveat that an unknown fraction may be AM artifact rather than biological pleiotropy

### Highest-leverage next steps for topology phase

1. **Build the trait correlation matrix**: Assemble published genetic correlations (from LD Score regression / GWAS) among the ~20–30 most well-characterized traits spanning cognition, personality, and psychopathology. Annotate each with AM-corrected estimates where available. This matrix is the empirical backbone of the topology.

2. **Map the developmental cascade**: Create a directed graph from temperament → personality → outcomes with age-indexed heritability and stability coefficients as edge weights. This captures the time dimension that a static correlation matrix misses.

3. **Formalize the variance decomposition**: For each major trait, create a standardized decomposition: [direct genetic] + [genetic nurture/indirect] + [AM-induced] + [shared environment] + [measured non-shared environment] + [stochastic residual]. Where values are unknown, flag them explicitly. This is the generating function skeleton that the formalization phase will flesh out.

---

## Model
*topic: human-psych-variation · stage: model · pass 8 · complete*

Generating function for human psychological variation. One equation per person; variance decomposition follows. Closed-form pieces: Crow–Felsenstein AM inflation, Wilson-Effect saturation, genetic-nurture additive split, multivariate sex-difference Mahalanobis D. Twin / SNP / within-family heritability are projections of the same decomposition. Interactive dashboard included.

## TLDR

The topology answered "what depends on what?". The formalization answers a sharper question: **given a person, where does their phenotype come from in expectation?** The answer is a single generating function that, once written down, dissolves several apparent paradoxes in the field — most importantly the gap between twin heritability, SNP heritability, and within-family heritability (they estimate different sums of the same underlying components, and the differences are informative).

The spine of this stage is one equation. Phenotype `P` for a person in a population is `P = A_d + A_i + A_LD + C + E_m + E_s + I`, with each term a contribution from a distinct mechanism: direct genetic effects from the person's own transmitted alleles, indirect genetic effects from parental (and broader-family) genomes operating through the environment they create, assortative-mating-induced linkage among causal variants, residual shared environment, measured non-shared environment, stochastic developmental noise, and gene-environment interaction terms. Variance decomposition follows directly, and is **block-orthogonal** rather than fully orthogonal: `V(P) = ΣV(component) + 2·Cov(A_d, A_i) + 2·Cov(A_d, E_m) + 2·Cov(A_d, C) + V(I)`. The cross-terms are the formal home of every gene-environment correlation finding in the literature; pretending they are zero is the most common modeling error. Three closed-form pieces drop out — the Crow–Felsenstein assortative-mating partition `V(A_LD) = h²_obs · r_δ` with `r_δ = m·h²_obs` (the dashboard *partitions* h² rather than inflating it; the equilibrium is reached in 5–10 generations of stable assortment), the Wilson-Effect logistic curve `h²(t) = h²_∞ / (1 + exp(−k·(t − t_50)))`, and the method gradient that says twin h² ≥ SNP h² ≥ within-family h² with the gaps decomposable into AM-LD, indirect-genetic, and rare-variant pieces.

A second module handles the multivariate sex-difference algebra, because the single largest framing trap in this field is the gap between univariate Cohen's d (typically 0.2–0.6 across personality dimensions) and the multivariate Mahalanobis distance `D² = Δμᵀ·Σ⁻¹·Δμ` (which can hit 2.7 when traits are weakly correlated and you stack 15 of them, as in Del Giudice 2012). The same data, two numbers, opposite-sounding stories — both correct. The formalization makes the bridge explicit so the reader can dial univariate d's and inter-trait correlations and watch D move.

What this stage **does not** formalize: the Plomin/Turkheimer interpretation of polygenic scores (verbal disagreement, no candidate equation), the mechanism behind the Gender Equality Paradox (three live hypotheses with no shared formalism), and the magnitude of AM-correction across the full cross-disorder genetic-correlation matrix (active research, methods just emerging). These remain at the observation stage; premature math here would mask uncertainty rather than reduce it. The L4 Lewontin firewall is preserved as a structural property of the model: the entire generating function is within-population, and nothing in it licenses between-population mean inference.

<PsychVariationModel client:load />

## How to read this stage

The dashboard above is the artifact. Everything below is the spec.

In plain language: when researchers report a "heritability of 0.50," what is being claimed is that *if you took the variance in a trait across a population and asked how much of it tracks genetic differences*, half of it does. It is a statement about the population's variance, not about any single person, and not about between-population differences. It says nothing causal beyond that — high heritability is fully compatible with large environmental effects (height is ~80% heritable and has risen ~10cm in a century). Different methods estimating "heritability" answer slightly different questions: twin studies pick up the broadest definition, within-family designs the narrowest. The gap between them is informative.

The stage formalizes that picture by writing one equation per person, decomposing it into named pieces, and showing how each measurement method projects onto a different subset of the pieces. Three closed-form sub-equations follow (assortative-mating inflation, Wilson-Effect age curve, genetic-nurture split). A second module addresses the same algebra applied to *group differences* — most prominently sex differences, but the framework is general. The dashboard lets you turn the knobs and watch the consequences.

You can read this top-down (TLDR → equation → closed forms → boundary conditions) or bottom-up (play with the dashboard, then come back to the equations when something surprises you). Either order works. The cruxes section at the end (§12) is where the load-bearing assumptions live; if any one fails, parts of the picture have to be rebuilt.

## 1. Move I'm making

This stage is a **decomposition + generating function + integration**, in that order:

- **Decomposition** — orthogonalize phenotypic variance into mechanism-specific components, with explicit non-orthogonal `Cov(G,E)` and interaction terms as the principled exceptions.
- **Generating function** — write the per-person phenotype as a deterministic function of those components plus stochastic noise. The variance decomposition follows by taking `V(·)` of the generating function.
- **Integration** — show that twin, SNP, and within-family heritability estimators are projections of the same underlying decomposition onto different observable subspaces. The Wilson Effect, AM inflation, and genetic-nurture findings then read as motion of those projections, not as separate phenomena.

What's **not** ready: anything in the topology marked O (open), and the polygenic-score causal-vs-summary debate, where the underlying disagreement isn't yet a formal one.

## 2. The generating function

For a single person `i` in a population at developmental time `t`, sampled from a stable mating regime:

```
P_i(t) = A_d,i + A_i,i + A_LD,i + C_i + E_m,i + E_s,i + I_i  +  μ(t)
```

| Term | Mechanism | Source identity |
|---|---|---|
| `A_d` | **Direct genetic** — additive effect of person's own transmitted causal alleles, evaluated as if mating were random | `Σ_k β_k · g_{ik}` over causal SNPs k |
| `A_i` | **Indirect genetic** (genetic nurture) — additive effect of parents' (and extended-family) genotypes operating through the rearing environment | parents' PGS × environmental transmission coefficient |
| `A_LD` | **Assortative-mating LD inflation** — additional additive variance induced by linkage among causal variants from non-random mating | At AM equilibrium, `V(A_d) + V(A_LD) = h²_obs`; the partition is `V(A_LD)/h²_obs = r_δ`. |
| `C` | **Shared environment residual** — environmental effects shared by siblings *not already captured* by `A_i`. Adult personality: ~0. Education / religiosity / politics: nonzero |
| `E_m` | **Measured non-shared environment** — identifiable causes (lead, schooling, head injury, peer composition, nutrition) | each enters with a measured causal coefficient, e.g. lead: β ≈ −6.2 IQ pts per 1–10 µg/dL |
| `E_s` | **Stochastic developmental noise** — unmeasured non-shared variance: developmental contingencies, immune/microbial, microscale neural variation, measurement error | the unmodeled residual; ~50% of personality variance |
| `I` | **Interaction terms** — `G×E`, `G×G` (epistasis), `G×age`. As of 2025 evidence, generally small at PGS-by-environment scale; large only at extreme environmental insults | residual non-additivity |
| `μ(t)` | **Population mean at age t** — not a person-level term but the developmental trajectory the person grows through | calibrated to age-norm tables |

**Why this form**: this is the additive-decomposition default of quantitative genetics extended with the two corrections that the 2018–2025 literature has installed into the field — separating `A_d` from `A_i` (Kong 2018, Young 2022) and separating `A_d` from `A_LD` (Border 2022, Yengo 2018, Wainschtein 2025). Earlier formulations folded `A_i` into `A_d` and `A_LD` into `A_d` and got the wrong answer about how much of the population-level genetic signal is direct biological causation. The within-family literature is what made these terms separately estimable.

**Scope note — scalar trait, not g-loaded vector**: `P_i(t)` is written as a scalar for one trait at a time. For cognitive ability, this collapses an underlying multi-ability structure (`g` + specific abilities, the CHC hierarchy) into a single phenotypic measure. The collapse is faithful when reporting `g`-loaded composite scores (e.g., full-scale IQ), and reasonable for any single primary ability. It is *not* faithful when the question is "how much of `A_d` for cognition is `g` versus specific abilities" — that requires a multivariate extension where each ability gets its own decomposition and `g` enters as a latent common factor across them. The topology's foundational assumption A3 (`g` exists as a real dimension) lives at this level: the model below operates *inside* a single ability/composite and inherits `g` as a property of which ability is being measured rather than as a structural component. For sex differences (Module B, §3.4), the multivariate extension is necessary by construction; that's why it appears as a separate module.

### 2.1 Variance decomposition

Taking variance of the generating function and tracking the cross-terms:

```
V(P) = V(A_d) + V(A_i) + V(A_LD)
     + V(C) + V(E_m) + V(E_s)
     + 2·Cov(A_d, A_i)        ← genetic nurture is correlated with direct effects (parents pass both)
     + 2·Cov(A_d, E_m)        ← active rGE: people select environments matching propensities
     + 2·Cov(A_d, C)           ← passive rGE residual (small once A_i is split out)
     + V(I)
```

The off-diagonal `Cov` terms are why "orthogonal decomposition" is the wrong frame for this system. The system is **block-orthogonal**: the additive components are roughly orthogonal *to the residual environment* but not *to each other*, and the cross-terms are the formal home of every gene-environment correlation finding in the literature. Pretending they're zero is the single most common modeling error.

### 2.2 Heritability identities

Three quantities are estimable from data; each picks up a different subset of the variance terms. The mapping is more subtle than a casual reading of the literature suggests, and it is worth getting right because the public-discourse confusion about "twin studies overestimate" turns on this exact algebra.

**The non-obvious point about twin h²**: `V(A_i)` (genetic nurture) is shared identically by MZ and DZ co-twins, because they share the same parents. Under a correctly specified ACE model, this variance lands in `C`, not `A`. So a faithful classical twin model **does not** count genetic nurture as heritability. The empirical observation that twin h² > within-family h² (e.g., for EA: 0.40 vs ~0.15) is therefore *not* due to twin h² capturing `A_i` directly. It is mostly due to two model-misspecification leakages: the ACE assumption rDZ_A = 0.5 fails under assortative mating (true sibling additive correlation under AM is `0.5·(1+r_δ)`), and the assumption that genetic nurture's contribution is fully shared between siblings can fail if parents differentially treat MZ vs DZ pairs.

| Estimator | What it estimates (correctly specified) | Practical leakages |
|---|---|---|
| **Twin h²** (classical ACE: `2·(rMZ − rDZ)`) | `V(A_d) + V(A_LD)` | Under unmodeled AM, some `V(A_i)` and `V(C)` bleed into A. Empirically, classical twin h² for EA exceeds within-family by ~0.20–0.25. |
| **SNP h²** (GREML, LDSC on population GWAS) | `V(A_d, common) + V(A_LD, common) + V(A_i, common)·attenuated` | Population GWAS effect sizes `β_pop = β_d + k·β_i` (where `k` is the AM coupling between transmitted and non-transmitted alleles), so SNP h² is **inflated by some `V(A_i)`**, but attenuated relative to the full `V(A_i)` because `k < 1`. Excludes rare variants. |
| **WGS h²** (Wainschtein 2025) | SNP h² + `V(A_d, rare) + V(A_LD, rare)` | Closes the rare-variant gap; same `A_i` contamination as SNP h² unless within-family. |
| **Within-family h²** (sib-FE, MZ-discordant, parent-offspring trios) | `V(A_d)` | Removes `A_i` and `A_LD` cleanly; leaves direct additive only. With WGS: `V(A_d) + V(A_d, rare)`. |

This is the **method gradient** (S2 in the topology):

```
twin h² ≥ WGS h² ≥ SNP h² ≥ within-family h²
```

The gaps are not measurement error. They are the data's way of telling you how much of "heritability" is structural (AM-LD), how much is environmental-via-parents (`A_i`), and how much depends on rare variants common-variant arrays cannot tag.

For educational attainment in 2025: classical twin h² ≈ 0.40, common-variant SNP h² ≈ 0.20–0.25, WGS h² ≈ 0.30 (with rare-variant contribution), within-family additive ≈ 0.15.

**Important caveat on Falconer's AM bias** (added in pass 6 after a reviewer correction). The "What it estimates" column above is exact only under random mating. Under positive AM, Falconer's formula `2·(rMZ − rDZ)` is biased *downward* by factor (1 − m_A) where m_A ≈ m·h² — fraternal twins share more than 50% of trait-relevant alleles because their parents are genetically more similar than chance, raising rDZ relative to rMZ and shrinking the formula's output. So Falconer estimates `[V(A_d) + V(A_LD)] · (1 − m_A) / V(P)`, not `V(A_AM)/V(P)` directly.

This matters for interpreting the gap. When classical twin h² > within-family h² for socially-structured traits (EA: 0.40 vs 0.15), the gap is dominated by *other* classical-ACE biases — primarily the equal-environments assumption (MZ co-twins are treated more similarly than DZ co-twins, inflating MZ correlation) and genetic-nurture leakage (V(A_i) leaks into A under model misspecification rather than landing cleanly in C) — *partially offset* by the AM downward bias. AM is a real phenomenon at the population level (Crow-Felsenstein V(A_LD) inflation; see §3.1 below) but it does not, on net, drive the twin-vs-within-family gap. The dominant inflation source is genetic nurture and EEA, with direct empirical anchors in Kong 2018 (non-transmitted PGS effect = 29.9% of transmitted for EA) and Okbay 2022 EA4 (within-family direct ~50% of population PGI). Within-family designs control for AM, EEA, and genetic nurture simultaneously.

This is the single calculation a careful reader of "twin studies vs molecular studies" headlines should be able to do. The numbers don't disagree; they answer different questions.

## 3. Closed-form pieces

Three components admit clean equations. The rest are calibrated empirically.

### 3.1 Assortative-mating inflation (Crow–Felsenstein)

There are two ways to use the AM-inflation formula, and they answer different questions.

**Forward problem** (rarely the relevant one): given the random-mating heritability `h²_rm` of a trait, what is the equilibrium heritability after stable AM? The answer is a fixed-point coupling `r_δ = m · h²*`, `V_A* = V_A / (1 − r_δ)`, `h²* = V_A* / (V_A* + V_E)`, reached in ~5–10 generations of stable assortment (Crow & Felsenstein 1968). One-iteration approximation: `r_δ ≈ m · h²_rm`, inflation `≈ 1 / (1 − m·h²_rm)`.

**Inverse / partition problem** (what the dashboard does): given the AM-equilibrium population additive variance `V(A_AM)/V(P) = h²_obs`, partition it into the random-mating-equivalent direct component `V(A_d)` and the AM-induced LD inflation `V(A_LD)`:

```
r_δ      ≈ m · h²_obs
V(A_d)   = h²_obs / (1 + r_δ + r_δ² + …)  =  h²_obs · (1 − r_δ)
V(A_LD)  = h²_obs − V(A_d)  =  h²_obs · r_δ
```

The Wilson curve gives `h²_obs(t)` directly, so the partition uses `r_δ = m · h²_obs(t)` with no iteration needed. This is what the dashboard implements. Pass-2 versions of the dashboard erroneously inflated `h²_obs` *on top of itself*, pushing twin h² above 1.0 at high parameter values; pass-4 corrected this.

**Note on what `h²_obs` should represent here** (added in pass 6 after a reviewer correction). The partition formula is a clean population-level decomposition of `V(A_AM)`. Different estimators recover `V(A_AM)/V(P)` with different biases: SNP-based heritability (GREML / LDSC on unrelated individuals) recovers it approximately unbiased; classical twin h² (Falconer) recovers `V(A_AM)/V(P) · (1 − m_A)` — biased downward by AM, partially offset upward by EEA violations and genetic-nurture leakage. The dashboard's Wilson-fit twin estimates conflate these biases. The partition formula's empirical validation is the match against SNP-based AM-LD estimates: Yengo 2018 measures `V(A_LD)/V(A)` for height at 14–23% empirically, matching the formula's prediction of `m·h² = 20%`. For socially-structured traits where Falconer twin h² is itself substantially inflated by EEA + genetic nurture, applying the partition formula to twin h² over-attributes the partition share to AM relative to its true population-level magnitude.

Worked anchors:
- **Educational attainment** with `m ≈ 0.4`, `h²_obs ≈ 0.40` (twin) → `r_δ ≈ 0.16`, `V(A_d) ≈ 0.34`, `V(A_LD) ≈ 0.06`. Caveat: the `h²_obs ≈ 0.40` here is the Falconer twin estimate, which is a biased proxy for the AM-equilibrium V(A)/V(P); applying the partition to the SNP-based estimate (~0.13) would give a smaller absolute V(A_LD).
- **Height** with `m ≈ 0.25`, `h²_obs ≈ 0.85` → `r_δ ≈ 0.21`, `V(A_d) ≈ 0.67`, `V(A_LD) ≈ 0.18` — matches the 14–23% empirical "AM-inflated" share Border et al. and Yengo et al. report (this is the trait where the partition's empirical validation is cleanest, because Falconer-bias vs SNP-h² discrepancies are smaller for height than for socially-structured traits).

Cross-trait AM (`m_xy ≠ 0`) extends the same logic to off-diagonal entries of the genetic-covariance matrix and is the formal reason E7 finds `R² = 0.74` between phenotypic-cross-mate correlations and genetic-correlation estimates. The cross-trait AM result (Border 2022) survives independently of the within-trait Falconer-bias issue: it's about between-trait LD inflating reported genetic correlations between disorders, which is empirically validated and not in dispute.

### 3.2 Wilson-Effect saturation curve

Heritability of cognitive ability rises with age because active rGE (G1) compounds: as children gain agency, they select environments matching their genetic propensities, amplifying genetic variance and shrinking shared environment. The empirical age curve from Bouchard 2013 and Briley & Tucker-Drob 2013 is sigmoidal — slow rise in early childhood, fastest gain in late childhood / early adolescence, saturation in late adolescence. A logistic gives a clean three-parameter fit:

```
h²(t) = h²_∞ / (1 + exp(−k_h · (t − t_50)))
```

With `h²_∞ ≈ 0.80`, `t_50 ≈ 9` years (age at half-asymptote), and `k_h ≈ 0.30/year`: `h²(5) ≈ 0.19`, `h²(10) ≈ 0.46`, `h²(15) ≈ 0.69`, `h²(25) ≈ 0.79`. These match Bouchard's anchors within ~3 percentage points across the full developmental range.

(Earlier passes used a saturating-exponential `h²_∞ − (h²_∞ − h²_0)·exp(−k·t)`, which rises too fast at the young end — it produced `h²(5) ≈ 0.52` for cognition vs the empirical ~0.20. The logistic form is the smallest functional change that fits the empirical sigmoidal pattern.)

The shared-environment trace runs an inverse path with a non-zero asymptote, since shared environment for cognition does not actually drop to zero in adulthood (~0.05 plateau is well-attested):

```
c²(t) = c²_∞ + (c²_0 − c²_∞) · exp(−k_c · t)
```

Cognition: `c²_0 ≈ 0.50`, `c²_∞ ≈ 0.05`, `k_c ≈ 0.15/year`. For Big Five personality, `c²_∞ ≈ 0` is appropriate (shared family environment effectively vanishes for personality by adulthood). For educational attainment and religiosity, `c²_∞ ≈ 0.10–0.15` should be substituted — these are exception traits where shared environment persists throughout life.

Both formulas are phenomenological — the parameters are not derived from a deeper model. They are calibration knobs for the dashboard.

### 3.3 Genetic-nurture decomposition (additive form)

Define `g_T` as the offspring's transmitted-allele PGS and `g_NT` as the parental non-transmitted-allele PGS. Then:

```
A_d = β_d · g_T
A_i = β_i · g_NT
```

Empirically (Kong 2018, Wang 2021, Okbay 2022, Howe 2022):

```
β_i / β_d ≈ 0.3 – 0.5  (educational attainment)
β_i / β_d ≈ 0.0 – 0.1  (height, BMI)
β_i / β_d ≈ 0.4 – 0.6  (cognitive performance)
```

**β-level vs variance-level.** The ratio `β_i/β_d` quoted above is at the *regression-coefficient* level. Translating to a variance contribution requires squaring (for the pure variance term) and an explicit cross-term:

```
V(A_i)              = β_i² · V(g)         =  (β_i/β_d)² · V(A_d)
2·Cov(A_d, A_i)     = 2·k · β_d · β_i · V(g)  =  2·k · (β_i/β_d) · V(A_d)
```

Where `k` is the AM-induced correlation between an offspring's transmitted-allele PGS and the parental non-transmitted-allele PGS. Under random mating `k ≈ 0` (Mendelian segregation makes them independent). Under stable AM, `k > 0` because spousal phenotypic correlation creates correlation between mom-transmitted alleles and dad-non-transmitted alleles (and vice versa); for AM-strong traits (EA, height) `k` is empirically in the 0.1–0.5 range, depending on the strength and stability of assortment.

This means `2·Cov(A_d, A_i)` is **not** generally larger than `V(A_i)`. For EA with `β_i/β_d ≈ 0.4`, `V(A_d) ≈ 0.15`, and a moderate `k ≈ 0.2`: `V(A_i) ≈ 0.024`, `2·Cov ≈ 2·0.2·0.4·0.15 = 0.024`. Total genetic-nurture variance contribution ≈ 0.048 — modest, on the same order as `V(A_i)` itself.

The dashboard displays only the pure `V(A_i) = (β_i/β_d)² · V(A_d)` slice as a clean variance bucket. The cross-term `2·Cov(A_d, A_i)` is the leakage path that makes empirical twin h² (under unmodeled AM) exceed the dashboard's "Twin h² (ACE)" output. It is acknowledged in the help text rather than allocated to a separate bar segment, partly because `k` is poorly constrained empirically and partly because adding a cross-term slice would over-clutter the visualization without changing the qualitative picture.

The relation `V(A_i) + 2·Cov(A_d, A_i) ≈ V_PGS,population − V_PGS,within-family` is approximate but useful — it turns "missing heritability after within-family correction" from a puzzle into an order-of-magnitude measurement. The exact RHS is `(β_i/β_d) · V(A_d) · (2k + (β_i/β_d))`, which depends on `k` and degrades to small values when AM is weak.

### 3.4 Multivariate sex-difference algebra (Module B)

For a trait vector `x` with covariance matrix `Σ` and group means `μ_F, μ_M`, the multivariate effect size is the Mahalanobis distance:

```
D² = (μ_F − μ_M)ᵀ · Σ⁻¹ · (μ_F − μ_M)
```

For uncorrelated traits with equal univariate effect sizes `|d|`, `D² = n·d²` so `D = d·√n`. For correlated traits, the inverse covariance structure either amplifies or shrinks `D` depending on whether sex-difference vectors are aligned with high-variance or low-variance directions of `Σ`.

**Worked example.** Take 15 personality dimensions (16PF), univariate `|d| ≈ 0.5` on average, with positive inter-trait correlations averaging `ρ ≈ 0.20`. Then approximately:

```
D² ≈ d² · 1ᵀ · Σ⁻¹ · 1
   ≈ d² · n / (1 + (n − 1)·ρ̄)        if Σ has a constant-correlation structure
   ≈ 0.25 · 15 / (1 + 14·0.20)
   ≈ 0.25 · 3.95
   ≈ 0.99
   D ≈ 1.0
```

The equicorrelated approximation undershoots Del Giudice 2012's reported `D = 2.71`, and this gap is informative rather than a bug. To recover 2.71 in the equicorrelated form would require average univariate `|d| ≈ 1.3`, far above what 16PF or NEO papers report at the observed level. What Del Giudice actually did was use **multigroup latent-variable modeling with measurement-error disattenuation**: he corrected each factor's `d` for unreliability and then computed `D` on the latent (true-score) means. Disattenuation magnifies effect sizes when reliability is well below 1.0, and aggregating across 15 factors then compounds the magnification. The honest summary is: at the level of *observed* (raw, pre-disattenuation) measurement, multivariate sex-difference `D` for personality is ~1.0–1.5; Del Giudice's 2.71 is the *latent-true-score* analogue.

The intuition behind the algebra still holds: if men and women differ on dimensions that are weakly correlated with each other, every dimension contributes independent information, and `D` grows with `√n`. If they differ on highly correlated dimensions, the differences carry redundant information and `D` plateaus. But the gap between observed and disattenuated `D` is itself a substantive piece of the field's debate — and worth flagging rather than papering over.

**Why this matters for distortions.** D3 (the "gender similarities" framing) cites univariate `d ≈ 0.05` for math performance and reads it as evidence of broad similarity. D4 (pop-evpsych framing) cites multivariate `D ≈ 2.71` and reads it as evidence of broad difference. Both citations are correct. The bridge equation shows that they are about different objects: a single dimension vs. a 15-dimensional space. Anyone who hasn't internalized this algebra can be silently captured by either framing.

**The algebra is general — not just for sex.** `D² = (μ_A − μ_B)ᵀ · Σ⁻¹ · (μ_A − μ_B)` applies to any two-group comparison: sex, age cohort, occupational sample, clinical vs. control, urban vs. rural — anywhere group means are reported on a multivariate panel. The module is presented in sex-difference language because that is where the framing trap concentrates, but readers thinking about other group comparisons can use the same dashboard. The L4 firewall (§5.2) does **not** block this generalization at the within-population level; it only blocks the leap from within-population variance/distance estimates to between-population *causal* claims. A descriptive `D` between two samples is fine; a *causal* interpretation of that `D` requires assumptions the model does not provide.

### 3.5 PGS portability decay (deferred)

Topology Variant C: `accuracy(distance)` calibration from Ding et al. 2023 (`r = −0.95` between genetic distance and PGS R² across 84 traits) is a clean candidate for closed-form. Deferred to a future tool because it sits at the population-genetics boundary rather than the within-population generative process this stage formalizes. Listed as a follow-up.

## 4. Composing the parts: anchors the dashboard preserves

The dashboard above stitches §3.1, §3.2, and §3.3 into one panel — sliders for trait class, age, `m`, `β_i/β_d`, and rare-variant share; outputs the variance decomposition and the three method-specific h² numbers. Four sanity-check anchors hold under the calibrated defaults:

1. **IQ at age 5 (cognitive)**: h²(5) ≈ 0.18, V(C) ≈ 0.26, V(A_d) ≈ 0.17. Matches Bouchard 2013.
2. **IQ at age 25 (cognitive, m=0.4, β_i/β_d=0.4)**: h²(25) ≈ 0.79, V(A_d) ≈ 0.54, V(A_LD) ≈ 0.25, V(A_i) ≈ 0.09, V(C) ≈ 0.06. Within-family h² (= V(A_d)) ≈ 0.54 — about a third more than the often-quoted EA within-family of 0.15, because cognition is a higher-h² trait than education.
3. **Big Five across adulthood**: h² ≈ 0.45, V(C) ≈ 0, V(E) ≈ 0.55. Effectively flat from age 5 onward.
4. **Variance budget closes**: V(A_d) + V(A_LD) + V(A_i) + V(C) + V(E) = 1.0 by construction. Twin h² never exceeds the Wilson asymptote.

These are the calibration targets. The biggest non-obvious one is anchor 4: the previous dashboard pass had the variance budget overflow under default parameters (twin h² > 1.0 at age 25 with m=0.4), which was a real bug. The current partition `h²_obs = V(A_d) + V(A_LD)` keeps the budget bounded by construction.

For traits the dashboard does not have a dedicated class for (educational attainment, height, religiosity, political affiliation), the user can approximate by choosing the closest class and adjusting sliders. EA-like behavior emerges from `cognitive` with m=0.4 and a mental note that h²(25) for EA is closer to 0.40 than 0.79 — i.e., the dashboard's cognitive class is calibrated to IQ, not EA.

## 5. Boundary conditions and where the model breaks

The generating function is correct only inside its scope. Five boundaries are explicit:

1. **Severe psychiatric tail.** The hyperpolygenic `A_d = Σ β_k g_{ik}` form assumes thousands of small effects. For early-onset autism with intellectual disability, single rare variants (CHD8, SCN2A) can carry effects of d > 1.0. The decomposition still works component-by-component but `A_d` becomes dominated by a small number of large-effect alleles — effectively Mendelian rather than polygenic. The model should either widen its prior on individual `β_k` or hand off to a separate Mendelian module at the tail.

2. **Between-population mean differences (L4 firewall).** Every term in the generating function is defined within a population at a stable mating regime. The model is structurally silent on between-population means: there is no `μ_pop` term to compare. Computing `D² = (μ_pop1 − μ_pop2)ᵀ Σ⁻¹ (μ_pop1 − μ_pop2)` is mathematically possible but requires assuming `Σ_pop1 = Σ_pop2` and equal causal architecture across populations — neither of which is empirically supported (Ding 2023's PGS-portability collapse is the empirical evidence that the assumption fails). This is the L4 / Lewontin firewall encoded directly into model scope.

3. **Severe environmental insults.** `V(I)` (interaction) is small at PGS-by-environment scale but large when environments cross threshold (lead, alcohol, severe deprivation, iodine). The additive decomposition under-fits at thresholds. Use the model in the normal range; switch to an explicit threshold-effect model at the extreme.

4. **Non-equilibrium AM.** The Crow–Felsenstein formula assumes AM has reached equilibrium. For populations under rapidly changing assortment regimes (e.g. rapid shifts in educational stratification), the inflation factor is en-route to the equilibrium value, not at it. Use the formula as an upper bound under those conditions.

5. **Individual-level inference (L1).** `V(A_d)` is a population variance. For a single person, `A_d` is a realization, not a partition. Statements like "70% of this individual's intelligence is genetic" do not type-check against the model. The dashboard exposes population variance only.

## 6. Distortion-aware reading

Each component of the decomposition has a public-discourse failure mode. The model's job is to make the failure visible, not to suppress it.

| Component | Common misreading | What the model says |
|---|---|---|
| `V(A_d)` (high) | "Genes determine outcomes" | Population variance. Says nothing about a specific person's prospects. |
| `V(A_i)` (large) | "Family environment doesn't matter" | The opposite: this term *is* family environment, mediated by parental genotypes that correlate with parental phenotypes. |
| `V(A_LD)` | Usually invisible to public discourse | Inflates V(A) at the population level by ~10–25% via AM-induced LD between trait-relevant alleles (Yengo 2018: 14–23% for height, matching the formula prediction). Does NOT on net inflate Falconer twin h² — AM actually biases Falconer downward, partially offsetting other classical-ACE biases (see §2.2 caveat). |
| `Cov(A_d, E_m)` (active rGE) | "People shape their environments" → therefore environments don't matter | They matter — the covariance term *is* their effect, just non-orthogonal to genes. |
| Twin h² ≥ within-family h² | "Twin studies overestimate" | They estimate a different quantity (population additive variance vs. direct effect). Both are real. |
| Multivariate `D` large | "Sexes are categorically different" | `D` is a distribution distance; individuals across the distributions still overlap substantially. Dimensional, not taxonic. |
| Univariate `d` small | "Sexes are essentially the same" | True for the dimension cited, false in the multivariate space. |

D1 and D2 (the two heaviest distortions) both operate by selecting a subset of these readings. The model doesn't resolve the political dispute, but anyone running the dashboard should be able to see why each side is technically correct about the term they're highlighting and incomplete about the rest.

## 7. Adversarial + steelman

Four objections to the formalization itself. The strongest version of each, then the model's honest response.

### Objection 1 — Variance bookkeeping is not a causal model

The decomposition partitions variance into named components, but it never specifies *why* `A_d` produces phenotype `P` rather than the reverse. A regression coefficient `β_d` from a within-family GWAS is not a causal effect; it is a statistical association under specific design assumptions. Calling the decomposition a "generating function" is false advertising — it generates *expected variance* given parameters, not *actual phenotype* given a causal mechanism.

**Steelman**: This is the strongest objection because it is the same disagreement that drives O1 (Plomin vs Turkheimer). The model accommodates both readings rather than picking one: under the Plomin reading, `β_d` is a causal coefficient and the decomposition is generative in the strong sense; under the Turkheimer reading, `β_d` is a regression coefficient that happens to be unbiased under within-family identification, but the underlying biology is unspecified. Both readings predict the same variance budget, which is why the data hasn't yet decided between them.

**Response**: Acknowledged. The model is more accurately described as a **conditional variance generating function** — given parameters, it generates the expected variance pattern. The causal interpretation of those parameters is exactly what's contested, and the decomposition's value is precisely that it lets both interpretations be expressed in a shared language. Stage 4 (data) is where the disagreement gets sharper: the test is whether `β_d, within-family` moves under environmental intervention. Plomin predicts no, Turkheimer predicts yes, and the model can express both predictions cleanly.

### Objection 2 — ACE assumptions are unrealistic enough that "twin h²" is not really estimating anything physical

EEA fails (MZ co-twins are treated more similarly than DZ); shared-environment effects vary by zygosity; non-shared environment for siblings is correlated with shared parental treatment. Stack the violations and the entire ACE framework is just a parametric reparameterization of the data, not an estimation of underlying components.

**Steelman**: Joseph and Richardson's critique of behavior genetics rests partly on this argument: the assumptions that make twin h² meaningful are violated enough to make the resulting numbers epistemically empty. The strongest version isn't that twin studies are "wrong" but that they're under-determined — multiple causal worlds produce the same `rMZ` and `rDZ` patterns.

**Response**: Partially conceded. Classical ACE is under-determined and the assumption-violation issue is real. However, two empirical findings constrain the under-determination: (a) **SNP-based heritability** (which uses unrelated individuals and bypasses EEA entirely) recovers a substantial fraction of twin h² across major traits — for height about 60% with common SNPs alone (rising to ~80% with whole-genome sequencing that captures rare variants), for cognitive ability about 25–40%, for educational attainment about 30–50%. The fraction varies by trait but is consistently non-trivial — the EEA-bias-only explanation for twin h² is empirically untenable; (b) **MZ-reared-apart studies** (Bouchard 1990 and updates) reproduce the basic Wilson Effect pattern with EEA structurally absent. The model takes twin h² as an *upper bound* on direct + indirect additive variance, not as a precise estimate. The method gradient is what makes the imprecision survivable: comparing twin to within-family bounds the gap.

### Objection 3 — Additive form misses dominance and epistasis

Dominance variance `V_D` and epistatic variance `V_I` (gene-gene interactions) are real and measurable. Twin studies fitting ADE models routinely find non-trivial `V_D`. The additive-only generating function is a simplification that loses information.

**Steelman**: For some traits — height (`V_D ≈ 0`), educational attainment (`V_D ≈ 0–0.05`) — dominance is small and the additive simplification is fine. For others — psychiatric disorders, where ADE models often outperform ACE — non-additive variance is potentially substantial. The additive-only model is not "wrong" so much as inappropriate for that subset of traits.

**Response**: Concede the scope limit. The generating function as written is for traits where polygenic-additive architecture dominates (which is most psychological traits, per Hill, Goddard & Visscher 2008's argument that even where dominance exists, additive variance often captures most of the variance because of allele-frequency distributions). For severe psychopathology and other traits with substantial `V_D`, an extended model would replace `A_d` with `A_d + D_d` and add `Cov(A_d, D_d)` cross-terms. The dashboard does not currently expose this; the prose acknowledges the boundary in §5.

### Objection 4 — Multivariate `D` conflates measurement structure with reality

Mahalanobis `D` depends on `Σ`. `Σ` is the within-sex covariance of measured traits, which depends on which traits you measure, how you measure them, and how the population varies. Different measurement panels produce different `D`s for the *same underlying difference*. The disattenuated `D = 2.71` from Del Giudice is not a property of human nature; it's a property of the 16PF + the U.S. sample + the latent-variable model.

**Steelman**: This is correct and underappreciated. `D` is not a population parameter in the way that `μ_F − μ_M` is. It is a model-relative summary statistic. Two researchers using different but equally defensible measurement panels can produce `D` values that differ by a factor of two or more.

**Response**: Conceded fully. The multivariate-D module's value is comparative, not absolute. It tells you *given a measurement structure*, how multivariate aggregation magnifies the apparent sex difference relative to any single dimension. The module's pedagogical purpose is to show why the same data (panel of dimensions) supports both "small per-dimension differences" and "large multivariate distance" — neither claim is wrong, but neither is the *whole* answer. The dashboard surfaces the dependency on `Σ` via the `ρ̄` slider so users can see how `D` moves under different correlation structures.

## 8. Open questions that the model exposes (Stage-4 inputs)

The formal apparatus makes four open questions sharper than verbal discussion alone:

- **O1 (PGS interpretation).** The decomposition treats `β_d · g_T` as a direct genetic term. Plomin's "PGS is a real biological cause" reading takes `β_d` as a structural causal coefficient. Turkheimer's "PGS is a summary of correlated environments" reading says `β_d` is contaminated by uncontrolled `Cov(A_d, E_m)`. The two interpretations make different predictions about how `β_d` should change under environmental intervention. **Stage 4 question**: for traits with large enough within-family GWAS, does `β_d, within-family` move under intervention (schooling reform, nutrition shifts) the way Plomin predicts (it shouldn't) or the way Turkheimer predicts (it should)?

- **O3 (Gender Equality Paradox).** The multivariate algebra in 3.4 shows that `D` depends on the inter-trait correlation structure `Σ`. If `Σ` differs between high-equality and low-equality societies, `D` will differ even if univariate `μ_F − μ_M` differences are fixed. **Stage 4 question**: does `Σ` (the personality covariance matrix itself) change across societies, or only the means? This is a different empirical question than "are the differences innate."

- **O6 (what `E_s` actually is).** The model treats stochastic developmental noise as an unmodeled residual. As Stage 4 data accumulates, candidates (immune/microbial, peer-network, epigenetic, measurement error) can be peeled off into `E_m` and the residual `E_s` should shrink. **Stage 4 question**: how much of the current ~50% personality `E_s` can be moved into `E_m` given current measurement panels?

- **O7 (cross-disorder rg post-AM correction).** Module 3.4's bridge between cross-trait phenotypic correlations and genetic correlations under AM (Border 2022, LAVA-Knock 2024) gives a formal correction. **Stage 4 question**: applied at scale to the full psychiatric-disorder rg matrix, what fraction of the cross-disorder genetic correlations survive the correction?

The two questions deferred from Section 1 (PGS portability and the GEP causal mechanism) are **not** sharpened by the model — they require new measurement, not new math.

## 9. Handoff to Stage 4 (data pipeline)

The model defines five parameter sets that Stage 4 needs to populate:

| Parameter | Source | Trait coverage |
|---|---|---|
| `β_d, β_i` | Within-family GWAS (Howe 2022, Okbay 2022) | EA, height, BMI, cognitive ability, depressive symptoms, smoking — extending |
| `m` (cross-spouse phenotypic correlation) | UK Biobank, HUNT, MoBa | EA, height, BMI, cognition, neuroticism — well-covered |
| `h²(t)` calibration | Bouchard 2013, Briley & Tucker-Drob 2013 longitudinal twin | Cognition (well-covered); personality (sparse); psychopathology (very sparse) |
| `Σ` for sex-difference module | Del Giudice 2012, Schmitt 2008, Kaiser 2020 | 16PF, NEO, Big Five |
| `share_rare` | Wainschtein 2025 | Height, EA, several psychiatric — extending |

The single highest-value Stage-4 deliverable: a per-trait table of `(twin h², SNP h², WGS h², within-family h², m, β_i/β_d)` at adulthood, ideally with cohort-by-age stratification. Most of the components already exist in published consortium summaries; the table is mostly aggregation, not new analysis.

## 10. Connection to adjacent topics

- **Parent-to-Child Transmission** (planned). The `A_i` term *is* the formal answer to "how much does parenting matter beyond genes for outcomes that look genetic." That topic should adopt this generating function as its starting point and refine `β_i` by domain (cognition vs. personality vs. health behaviors) and by mechanism (vocabulary input, expectation-setting, neighborhood selection). The Nivard et al. 2024 finding — that indirect genetic effects extend beyond the nuclear family — implies `β_i` should be further decomposed into a parent-level term and a dynastic/extended-family term.

- **Evolution-Modernity Mismatch** (planned). The `μ(t)` population-mean trajectory is the formal home of secular shifts (Flynn rise, Flynn reversal, age-of-puberty drift). Within-cohort within-sibship designs are the cleanest separator of genuine environmental shifts in `μ(t)` from compositional or selection artifacts. The Pietschnig 2024 finding that the positive manifold itself may be weakening across recent cohorts suggests `μ(t)` is not a one-dimensional curve but a moving structure of which abilities are gaining or losing — which the current scalar form does not capture.

(A connection to a planned "Bedrock Generating Functions" topic was floated in pass 1 but dropped — the analogy was real but too loose to do useful work here, and any cross-domain claim should live in that topic's own formalization rather than be asserted from this one.)

## 11. Glossary (formalization-specific additions)

This section's symbols are listed in the order they appear in the generating function. The lit-review and topology glossaries cover the field-level terminology (h², SNP, GWAS, PGS, AM, rGE, GxE, etc.) and are not duplicated here.

| Symbol / term | Meaning |
|---|---|
| `P_i(t)` | Phenotype of person `i` at developmental age `t`. Scalar (for one trait at a time); see §2 scope note for the multi-ability extension. |
| `A_d` | **Direct additive genetic component** — `Σ_k β_k · g_{ik}` over causal SNPs the person inherits, evaluated as if mating were random. |
| `A_i` | **Indirect additive genetic component** (genetic nurture) — additive effect of parents' (and extended-family) genotypes operating through the rearing environment. |
| `A_LD` | **AM-induced LD inflation** — additional additive variance from non-random mating creating linkage among causal variants. |
| `C` | Shared-environment residual not already absorbed by `A_i`. |
| `E_m` / `E_s` | Measured non-shared environment (lead, schooling, etc.) / stochastic developmental noise (the unmodeled residual). |
| `I` | Interaction terms: `G×E`, `G×G` (epistasis), `G×age`. |
| `μ(t)` | Population-mean trajectory at age `t` (developmental norm, not a person-level term). |
| `β_d` / `β_i` | Direct / indirect genetic regression coefficients on phenotype, estimated from within-family / parental-genotype designs. |
| `g_T` / `g_NT` | Polygenic score from offspring's transmitted alleles / parents' non-transmitted alleles. |
| `m` | Cross-spouse phenotypic correlation (assortative-mating strength on the measured trait). |
| `r_δ` | Cross-spouse correlation in additive genetic value; `= m · h²_obs` at AM equilibrium. The dashboard uses this directly (no fixed-point iteration) since Wilson h²(t) is already the equilibrium quantity. |
| `k` | AM-induced correlation between transmitted and non-transmitted alleles within parents; appears in the genetic-nurture variance identity (§3.3). |
| `V_A*` | Additive genetic variance at AM equilibrium; `V_A* = V_A / (1 − r_δ)` in the Crow–Felsenstein form. The dashboard observes V_A* directly via h²_obs and uses the formula to *partition* it into V(A_d) and V(A_LD). |
| `h²(t)` | Heritability as a function of age; logistic form `h²_∞ / (1 + exp(−k·(t − t_50)))` in §3.2. Earlier passes used a saturating exponential which fit the asymptote but rose too fast in childhood; the logistic is the smallest functional change that captures the empirical sigmoidal pattern. |
| `Σ` | Trait-level (within-sex or within-group) covariance matrix used in multivariate-D calculation. |
| `Mahalanobis D` | Multivariate generalization of Cohen's d: `√(Δμᵀ Σ⁻¹ Δμ)`. |
| `ρ̄` | Average inter-trait correlation in `Σ`; the equicorrelated approximation collapses `D²` to `d²·n/(1+(n−1)ρ̄)` (§3.4). |
| `block-orthogonal` | Decomposition where major components are orthogonal to the residual environment but cross-terms within components (e.g. `Cov(A_d, A_i)`) are explicit, not zero. |
| `method gradient` | The relationship `twin h² ≥ WGS h² ≥ SNP h² ≥ within-family h²` driven by which components each estimator includes. |

## 12. Cruxes for this model

The topology had cruxes for the field. This stage's cruxes are different — they are the load-bearing assumptions of the *formalization* itself. If any one flips, the model needs to be restructured.

| Crux | Load-bearing claim | What would flip it |
|---|---|---|
| **C1** | Within-family GWAS effect estimates are an unbiased estimate of `β_d`. The whole `A_d` / `A_i` separation depends on this. | A demonstration that within-family designs have a systematic confound (e.g., differential parental treatment that correlates with offspring genotype) that biases `β_d` by more than ~10%. So far Howe 2022 / Okbay 2022 within-sibship GWAS are mutually consistent and consistent with trio-based estimates, suggesting the confound is bounded. |
| **C2** | AM equilibrium has been reached or is close enough that the partition relation `r_δ = m·h²_obs` holds. | A demonstration that recent population-scale shifts in assortment (educational stratification expansion since 1970, online dating since 2010) have moved populations far from equilibrium for psychologically-relevant traits — at which point the observed `r_δ` would lag the formula's prediction. Currently no direct evidence the partition is mis-calibrated; would require longitudinal `m`-by-cohort data. |
| **C3** | Hyperpolygenic architecture: `A_d = Σ β_k g_{ik}` over thousands of small effects, no single locus dominates. | Discovery that for a major psychological trait class, ~5–10 large-effect variants account for >50% of `V(A_d)`. Currently true only for the severe psychiatric tail (autism with ID, severe schizophrenia spectrum), where the model already concedes scope (§5.1). Would generalize to mainstream cognition only if a CRISPR-era discovery overturned the polygenic consensus. |
| **C4** | `A_d`, `A_i`, `A_LD` are jointly identifiable given the available designs. | A demonstration that twin/SNP/within-family/WGS estimators are *not* sufficient to disentangle all three (e.g., that AM-LD and rare-variant contributions are mutually confounded in a way no current design can break). This would force collapsing the decomposition or treating one component as a residual. Active concern: rare-variant heritability in WGS may itself be inflated by AM-LD among rare variants, which would muddy C4. |
| **C5** | Equicorrelated `Σ` is a useful approximation for the multivariate sex-difference module. | A demonstration that real personality covariance matrices have block-structured (or low-rank) `Σ` that produces qualitatively different `D` from the equicorrelated approximation. Already partially true: 16PF has known higher-order factor structure, which is why the equicorrelated approximation undershoots Del Giudice's latent-variable result. Crux holds in a weakened form: equicorrelated is useful pedagogically but not quantitatively for high-dimensional panels. |

The most consequential of these is **C4**. If `A_d`, `A_i`, and `A_LD` cannot be jointly identified by current designs, the variance decomposition reduces to a coarser partition (genetic-additive vs everything else), and the field-level dispute about how much "genetic" effect is environment-mediated remains parametrically unresolvable rather than just empirically pending.

---

## Topology
*topic: human-psych-variation · stage: topology · pass 4 · complete*

Dependency graph of the lit review. Three categories of high-stakes node (foundational cruxes / reframer nodes / logical guardrails) plus weakest links, four variant views, three Stage-3 options, an objections section, and a glossary. Updated through 2024-2025 literature on AM correction, within-family GWAS at scale, PGS portability, Scarr-Rowe collapse, GEP replication, and missing-heritability closure.

## TLDR

The lit review documents what the science says about psychological variation. This topology asks a sharper question: **what depends on what?** Strip the field down to its load-bearing structure and the picture is surprisingly clean. Three foundational assumptions — that twin/adoption methods give approximately valid variance decomposition (A1), that GWAS signal reflects real genetic effects rather than population structure or assortative-mating artifact (A2), and that a general factor of cognitive ability `g` exists as a real dimension of individual variation (A3) — sit upstream of most of the empirical and synthesis nodes in the graph; if any one of them flipped, large regions would have to be rebuilt. Everything else is either an empirical claim resting on these foundations, a methodological prerequisite that lets the foundations be tested, a logical necessity that constrains how the empirical claims can be interpreted, or a generating mechanism that explains why the empirical pattern looks the way it does.

The high-stakes nodes split into three categories — keeping them separate is the single most useful conceptual move in this topology. **Foundational cruxes** (A1 twin validity, A2 GWAS signal real, A3 g exists) are the assumptions that, if falsified, force rebuilding regions of the picture. **Reframer nodes** (G2/E6 passive rGE / genetic nurture; G6/E7 cross-trait assortative mating) don't break the picture if reversed — they change what it *means*; their magnitudes are being actively quantified and their precise share of population-level "genetic" effects is the field's most consequential open quantity. **Logical guardrails** (L1 variance-ratio definition; L4 Lewontin firewall) cannot be falsified — they can only be ignored, which is exactly how most public-discourse misuse of the field proceeds. Conflating these three types under a single label of "important findings" is a major source of bad-faith debate.

The field's **weakest links** are not where public discourse focuses heat. Mainstream contests over "is heritability real" target settled findings (A1+E1 are robust); the 2025 whole-genome-sequencing work (Wainschtein et al. 2025, *Nature*) now closes ~88% of the pedigree-based heritability gap, so the "missing heritability" critique is also substantially answered. The actual fragile zones in 2026 are: (a) the generalization from candidate-GxE failure to all-GxE-is-small — partially holding the null (Allegrini 2020 for education, 2025 systematic review for depression) but the literature is still too young for confidence; (b) **Scarr-Rowe** has weakened further — Ghirardi et al. 2024 found 39/42 PGI×SES interactions in the *opposite* (compensatory) direction, so "deprivation suppresses heritability" is now evidence-thin; (c) the polygenic-score → mechanism inference (Plomin's "causal" view vs. Turkheimer's "weak explanation" view) remains genuinely undecided; (d) the magnitude of AM-correction across psychiatric cross-disorder rg estimates is now being actively addressed — Ma, Wang, Border et al. 2024 (LAVA-Knock) is the first method to systematically reduce xAM-induced bias, with the field-wide answer likely in 2–3 years. The Flynn-reversal cause and the Gender Equality Paradox mechanism remain open mechanistic questions, but they are open in a different way — the empirical patterns themselves are robust; only the explanation is contested. The 2025 GEP systematic review (Herlitz et al.) actually *strengthened* the pattern across personality, verbal abilities, episodic memory, and negative emotions.

This topology is the input to model formalization (Stage 3). The cleanest formalization target is the **variance decomposition equation**: V(trait) = V(direct genetic) + V(genetic nurture / indirect) + V(AM-induced LD) + V(shared environment) + V(measured non-shared) + V(stochastic) + 2·Cov(genes, environment) + interaction terms — with each term parameterized by trait, age, and population context, and with the AM and rGE terms being where current methodological revision is concentrated. The four variant views below (Vulnerability / Flow / Minimal / Politicization) read the same graph through different lenses to make the formalization choices easier.


## The graph

<PsychVariationGraph client:load />

*Click a node for its claim and load-bearing weight; hover an edge for the relation type; drag to rearrange. The variant toggles read the same graph through different lenses (vulnerability, flow, minimal claim set, politicization).*

---

## How to read this graph

Every node in the lit review collapses to one of eight types. Edges between them carry one of seven relations. Together they make the structure inspectable.

### Node types

| Code | Type | What it is |
|---|---|---|
| **A** | Foundational assumption | A claim the field cannot operate without; if false, large downstream regions collapse |
| **M** | Methodological prerequisite | A study design or estimation tool that must work for the empirical claims to be testable |
| **E** | Empirical claim | A specific measured finding with an effect size and replication status |
| **L** | Logical necessity | Follows from definitions or algebra; not empirically refutable |
| **G** | Generating mechanism | A causal process that explains a pattern (rGE, AM, niche-picking, critical periods) |
| **S** | Synthesis claim | An integrative statement combining multiple lower-level claims |
| **O** | Open question | Genuinely undecided with current methods or evidence |
| **D** | Distortion vector | Where motivated reasoning concentrates (typed by direction) |

### Edge types

| Code | Edge | Meaning |
|---|---|---|
| **dep** | depends-on | If target collapses, source collapses |
| **imp** | implies | Logical implication |
| **sup** | empirically-supports | Evidence relation |
| **conf** | confounds / inflates | Artifact relationship (e.g., AM inflates rg) |
| **mod** | moderates | Changes magnitude (e.g., SES × heritability) |
| **dev** | develops-into | Temporal/developmental successor (temperament → personality) |
| **corr** | corrects | Within-family corrects between-family bias |

### Weight scale (load-bearing weight, 1–5)

- **5** — crux node; collapse propagates across multiple sections of the lit review
- **4** — load-bearing within a section
- **3** — important but local
- **2** — corroborating
- **1** — decorative; could be removed without changing the picture

---

## 1. Node catalog

Each node carries: type code · weight · short claim · key citation · status. Status flags: ✓ (robust/replicated), ~ (partial/qualified), ? (contested), ✗ (refuted, kept as historical reference).

### A — Foundational assumptions

| ID | Wt | Claim | Status |
|---|---|---|---|
| **A1** | 5 | Twin/adoption methods provide approximately valid variance decomposition (EEA modestly violated but not fatally) | ✓ |
| **A2** | 5 | GWAS signal reflects real genetic effects, not (only) population stratification or AM artifact | ✓ partial |
| **A3** | 5 | A general factor of cognitive ability `g` is a real dimension of individual variation (positive manifold) | ✓ statistical / ? mechanism |
| **A4** | 3 | Heritability findings apply to the population sampled, not to individuals or other populations (scope) | ✓ |
| **A5** | 3 | Phenotypes are reliably and validly measurable across cultures and time | ~ |
| **A6** | 3 | Most psychological variation is dimensional, not taxonic | ✓ |

### M — Methodological prerequisites

| ID | Wt | Tool | Notes |
|---|---|---|---|
| **M1** | 4 | Twin studies (MZ/DZ) at scale | Polderman 2015 meta: 14.5M pairs |
| **M2** | 4 | Adoption studies, especially cross-cultural (Korean-American) | Sacerdote 2007; Beauchamp 2023 |
| **M3** | 5 | GWAS at N ≥ 100k (ideally ≥ 1M for personality/EA) | Okbay 2022 (3M for EA) |
| **M4** | 5 | Within-family designs (sibling FE, MZ-discordant, parent-offspring trios). Kong 2018; Okbay 2022 (EA, N=3M); Howe et al. 2022 *Nature Genetics* extended this to 178k siblings × 25 phenotypes — within-sibship estimates were systematically smaller than population estimates for height, EA, cognitive ability, depressive symptoms, smoking. The within-family approach is now mature beyond just educational attainment | ✓ |
| **M5** | 4 | Polygenic scores (PGS) | Best R² ~0.16 for EA, ~0.10 for SCZ |
| **M6** | 3 | Cross-trait LD-score regression for genetic correlations | Brainstorm 2018 |
| **M7** | 3 | Mendelian randomization | For causal inference from observational data |
| **M8** | 3 | Pre-registration & collaborative meta-analysis | Demolished candidate-GxE |
| **M9** | 4 | Whole-genome sequencing (rare-variant capture). 2025 follow-up (UK Biobank ~500k, Wainschtein et al. 2025 *Nature*) captures ~88% of pedigree-based narrow-sense heritability across many traits (20% rare + 68% common variants). The "missing heritability" problem is now substantially resolved for many phenotypes | ✓ |

### E — Empirical claims

Cognition / IQ:
| ID | Wt | Claim | Status |
|---|---|---|---|
| **E1** | 5 | Mean trait heritability ≈ 0.49 across 17,804 traits (Polderman 2015) | ✓ |
| **E2** | 5 | Shared environment C ≈ 0 for adult personality and most adult cognition | ✓ with exceptions (EA, religiosity, politics) |
| **E3** | 4 | Wilson Effect: IQ heritability rises from ~0.20 (age 5) to ~0.80 (adulthood) | ✓ |
| **E4** | 5 | Hyper-polygenic architecture: thousands of small-effect variants | ✓ (Turkheimer's 4th Law, Chabris 2015) |
| **E5** | 4 | Candidate-gene approach for psychiatric/personality traits failed (5-HTTLPR etc.) | ✗ original claims; ✓ collapse finding |
| **E6** | 5 | Within-family PGS effects are ~½ population-level effects (genetic nurture) | ✓ for EA, BMI, height |
| **E7** | 5 | Cross-trait assortative mating explains R²≈74% of variance in genetic correlation estimates | ✓ (Border 2022, Science) |
| **E8** | 4 | Lead exposure 1–10 µg/dL → −6.2 IQ pts (Lanphear 2005) | ✓ |
| **E9** | 4 | Each year of schooling adds ~3.4 IQ points, persisting into old age | ✓ (Ritchie & Tucker-Drob 2018) |
| **E22** | 4 | Within-population heritability does not license between-population inference | ✓ logical |
| **E23** | 4 | PGS prediction accuracy decays *continuously* along the genetic-distance continuum from training population (Pearson r = −0.95 between genetic distance and PGS accuracy across 84 traits, Ding et al. 2023, *Nature*). Reframes the older "discrete ancestry-group drop" picture (Martin 2019; Mostafavi 2020) | ✓ |
| **E24** | 3 | Flynn Effect and its post-1990s reversal are both environmentally driven (within-sibship evidence) | ✓ pattern; ? cause |
| **E25** | 2 | Scarr-Rowe: SES × heritability hypothesis (more genetic expression in higher-SES). Weakening further in 2024 — Ghirardi et al. 2024 found 39/42 PGI×SES interactions in education NEGATIVE (compensatory direction); only 1 significant positive. Pattern is now closer to "compensatory hypothesis holds, Scarr-Rowe fails" than to "context-dependent" | ✗ |
| **E18** | 4 | Positive manifold: every cognitive test correlates positively with every other | ✓ |
| **E26** | 3 | Childhood IQ → all-cause mortality: each 1-SD ≈ 24% lower mortality | ✓ (Calvin 2011) |
| **E27** | 3 | Lifespan IQ stability: Lothian Birth Cohort age-11 → age-90 r ≈ 0.67 | ✓ |
| **E28** | 3 | Severe iodine/alcohol/deprivation cause large asymmetric IQ effects | ✓ |

Personality / temperament:
| ID | Wt | Claim | Status |
|---|---|---|---|
| **E29** | 4 | Big Five h² ≈ 0.40–0.60; cross-cultural replication for E/A/C | ✓ partial (Tsimane qualifier) |
| **E30** | 4 | Cumulative continuity: rank-order stability rises to ~0.74 by midlife | ✓ |
| **E31** | 4 | Maturity principle: mean-level ↑ in C, A, ES with age | ✓ |
| **E32** | 3 | Temperament dimensions (Surgency, Negative Affectivity, Effortful Control) → adult personality | ✓ |
| **E33** | 3 | Personality predicts mortality/divorce/income at magnitudes ≈ SES & cognition | ✓ (Roberts 2007) |

Sex differences:
| ID | Wt | Claim | Status |
|---|---|---|---|
| **E10** | 4 | Multivariate Big Five sex difference: D ≈ 2.71 (~10% overlap) | ~ method-sensitive |
| **E11** | 4 | People-things interest difference d ≈ 0.93 (largest in psychology) | ✓ |
| **E12** | 3 | Mental rotation d ≈ 0.56–0.73 male advantage | ✓ |
| **E13** | 3 | Math performance d ≈ 0.05–0.10 (essentially equal) | ✓ |
| **E14** | 4 | Gender Equality Paradox: differences larger in egalitarian/wealthier countries. Herlitz et al. 2025 systematic review (54 articles, 27 meta-analyses) confirmed the pattern across personality, verbal abilities, episodic memory, and negative emotions — pattern replication has strengthened, not weakened | ✓ pattern; ? mechanism |
| **E34** | 3 | Physical aggression d ≈ 0.40–0.60 male; ~95% of homicides male | ✓ |
| **E35** | 2 | CAH girls show masculinized toy preferences; primate parallels | ✓ |

Psychopathology:
| ID | Wt | Claim | Status |
|---|---|---|---|
| **E15** | 4 | All major psychiatric disorders highly heritable (h² 0.35–0.85) and hyper-polygenic | ✓ |
| **E16** | 4 | Cross-disorder genetic correlations exist | ✓ existence; ? magnitude post-AM |
| **E17** | 3 | A `p` factor (general psychopathology) fits cross-syndrome data | ✓ statistical; ? interpretation |
| **E36** | 3 | Autism: common-PGS positively correlated with IQ; rare/de-novo drives ID-comorbid cases | ✓ |
| **E37** | 2 | Critical-period plasticity (GABAergic, perineuronal nets) is mechanistically real | ✓ |

### L — Logical necessities

| ID | Wt | Claim |
|---|---|---|
| **L1** | 5 | Heritability is a population variance ratio; it does not partition individual phenotypes (mathematical form — A4 is the scope-of-claim sibling) |
| **L2** | 4 | h² changes with environmental variance: hold genes constant, equalize environments → h² → 1 |
| **L3** | 5 | Within-family designs control for between-family confounds (rGE, stratification, AM) |
| **L4** | 5 | Within-population heritability provides no information about between-population mean differences (Lewontin). E22 in the empirical column is the applied form of this same point |
| **L5** | 3 | Multivariate D ≥ max(univariate d) when component dimensions are positively correlated |
| **L6** | 3 | Positive manifold permits both unitary-cause and emergent-network interpretations of g |
| **L7** | 3 | Effect-size interpretation is scale-dependent (d=0.10 trivial in trait psychology, large in clinical) |

### G — Generating mechanisms

| ID | Wt | Mechanism | Drives |
|---|---|---|---|
| **G1** | 4 | Active rGE / niche-picking | E3 (Wilson Effect amplification) |
| **G2** | 5 | Passive rGE. Wang 2021 / Isungset 2022 confirm indirect ≈ ½ direct genetic effect for EA. Nivard et al. 2024 found indirect genetic effects on offspring achievement extend *beyond* the nuclear family — dynastic / extended-family / community processes contribute, so the "parents transmit gene + correlated environment" framing understates the spread | E6 (genetic nurture); inflation of population-level h² |
| **G3** | 3 | Evocative rGE | Heritability of "environments" (Kendler & Baker 2007) |
| **G4** | 3 | Critical-period plasticity (GABAergic maturation) | Asymmetric environmental effects on early development |
| **G5** | 4 | Assortative mating → LD induction | Inflates additive genetic variance V(A) at the population level (Yengo 2018: 14–23% for height); inflates SNP h² and PGS effect sizes. Counterintuitively biases Falconer twin h² downward (raises rDZ relative to rMZ); twin-vs-WF gap for socially-structured traits is dominated by genetic nurture / EEA, with AM partially offsetting. |
| **G6** | 5 | Cross-trait AM → spurious genetic correlations | Confounds E16, E17 (p-factor) |
| **G7** | 4 | Stochastic developmental noise | Dominant source of non-shared environment |
| **G8** | 3 | Selection / niche construction across the lifespan | Bridges temperament → personality → outcome cascade |

### S — Synthesis claims

| ID | Wt | Claim |
|---|---|---|
| **S1** | 5 | "Genes vs. environment" is the wrong frame; the system is tightly coupled (genome × rGE × AM × few large environmental insults × stochastic noise × culture × developmental unfolding) |
| **S2** | 5 | Twin h² ≥ SNP h² ≥ within-family h² gradient quantifies AM/rGE/measurement inflation across estimation methods |
| **S3** | 4 | Heritability ≠ destiny; high h² is compatible with large environmental shifts (height: h² ≈ 0.80, +10cm in a century) |
| **S4** | 4 | Most "non-shared environment" is stochastic, not systematic — it accounts for ~50% of personality variance and is poorly characterized |
| **S5** | 4 | Two parallel hierarchies (CHC for cognition, HiTOP for psychopathology) connected at the top by inverse g↔p genetic correlation |
| **S6** | 4 | Developmental cascade: temperament (infant biological reactivity) → personality (adult social-cognitive layer added) → outcomes (mortality, attainment, relationships) with h² ↑ and shared-env ↓ across the lifespan |

### O — Open questions

| ID | Wt | Question | Why it matters |
|---|---|---|---|
| **O1** | 5 | Mechanistic interpretation of PGS: "causal genetic" (Plomin) vs. "weak explanation" (Turkheimer) | Determines what PGS prediction *means* |
| **O2** | 3 | Cause of Flynn-effect reversal post-1990s | Empirical pattern robust (Bratsberg & Rogeberg 2018 within-sibship Norway). Mechanism still unsettled. Pietschnig et al. 2024 (Vienna 2005–2018 cohort) added a wrinkle: the *positive manifold itself* may be weakening — gains in some abilities aren't tracking gains in others, suggesting the g-loading of the rise/fall is not constant. Hypothesized mechanisms (screens, reduced long-form reading, attention) circulate without empirical pinning |
| **O3** | 4 | Causal mechanism behind Gender Equality Paradox | Innate-expression release vs. measurement artifact vs. confound — selection of explanation has political stakes |
| **O4** | 5 | Between-population mean differences: any genetic component? | **Currently scientifically unanswerable** with available methods (PGS portability too poor, cross-ancestry GWAS at scale don't exist). Honest position: unresolved, not settled in either direction |
| **O5** | 3 | g architecture: latent common cause vs. emergent network (mutualism, van der Maas 2006) | Affects how interventions could in principle move g |
| **O6** | 4 | What "non-shared environment" actually is: stochastic noise, immune/microbial, peer networks, epigenetic, measurement error | Largest unmodeled variance component in personality |
| **O7** | 5 | Magnitude of AM-correction across the cross-disorder genetic correlation matrix | Active revision. Ma, Wang, Border et al. 2024 *AJHG* introduced LAVA-Knock — a local-genetic-correlation method that reduces xAM-induced bias. Methods to give the answer are now emerging, not just to flag the problem |

### D — Distortion vectors (where motivated reasoning concentrates)

| ID | Direction | Targets | Failure mode |
|---|---|---|---|
| **D1** | Blank-slate / environmentalist | A1, E1, E10–E14 | Dismiss twin studies wholesale; oversell transgenerational epigenetics; overstate stereotype threat; minimize sex differences via univariate-only framing |
| **D2** | Hereditarian | L4, E22, E23, O4 | Ignore Lewontin; treat g-loadedness of gaps as evidence of genetic etiology; cite fringe admixture studies; ignore AM/rGE corrections to PGS |
| **D3** | "Gender similarities" minimization | E10, E11, E14 | Selective citation (math d=0.05) to imply no differences anywhere; obscure D=2.71 multivariate; minimize d=0.93 interest gap |
| **D4** | Pop evpsych overgeneralization | E10–E14, A6 | Treat dimensional ds as taxonic; extrapolate small ds to categorical claims; overgeneralize from specific tasks to broad-domain claims |

---

## 2. Dependency cascade

The cascade reads from foundations up to synthesis, and from corrections back down to corrected claims.

### Forward cascade (foundations → empirical claims → synthesis)

```
A1 ──dep──> M1, M2 ──sup──> E1, E2, E3, E25, E29
A2 ──dep──> M3, M5 ──sup──> E4, E6, E7, E15–E17, E22–E23
A3 ──dep──> E18 ──sup──> E26, E27 ──imp──> S5
A4 (scope) + L1 (form) ──guards──> interpretation of E1, E2, E3 and S3
A6 ──imp──> S5 (dimensional turn in psychiatry)
M9 ──corr──> E1 (closes missing-heritability gap)
M3 + M4 ──sup──> E6 (genetic nurture), E7 (xAM)
E5 (candidate-gene collapse) ──imp──> E4 (polygenic architecture confirmed by absence of large hits)

E1 + E2 + E3 + G1 ──imp──> S6 (developmental cascade)
E1 + E4 + G2 + G5 ──imp──> S2 (h² gradient by method)
E10 + E11 + E12 + E13 + L5 ──imp──> "small univariate, large multivariate" sex-difference picture
E14 + O3 ──imp──> mechanism-pending GEP
E15 + E16 + E17 ──imp──> S5 (HiTOP/p)
E22 + E23 + L4 ──imp──> O4 (between-pop unanswerable currently)
E1 + E2 + E4 + E6 + E7 + E8 + E9 + G1–G7 ──imp──> S1, S2 (integrated picture)
```

### Backward / corrective cascade (newer evidence revises older claims)

```
G2 (passive rGE) ──corr──> E1 estimates (population-level overstates direct genetic)
G6 (cross-trait AM) ──corr──> E16 (some psychiatric rg's may be xAM artifact)
M4 (within-family) ──corr──> E6 magnitude (~½ of population PGS)
M8 (preregistration) ──corr──> E5 (collapsed candidate-GxE)
M9 (WGS) ──corr──> "missing heritability" interpretation
```

### Distortion → target edges

```
D1 ──attacks──> A1, E1, E10, E11, E12, E14
D2 ──attacks──> L4, E23 (ignores), exploits A2 absent corrections from G2/G6
D3 ──attacks──> E10, E11 (selective univariate framing)
D4 ──attacks──> A6, L7
```

---

## 3. Where pressure concentrates

A common failure mode in this literature is to treat all high-stakes nodes as the same type of thing. They are not. The graph has three distinct categories of high-stakes node and one category of fragile claim — keeping these separate sharpens what the field actually needs to resolve.

### 3a. Foundational cruxes — falsification breaks regions of the picture

These are the empirical-or-methodological assumptions that, if wrong, force rebuilding large parts of the lit review.

**A1 — Twin/adoption method validity.** Carries Section 1 of the lit review; heritability-by-domain table; Wilson Effect. Robustness: HIGH (MZ-reared-apart, SNP-h² bypassing EEA, misperceived-zygosity all converge). Would flip if SNP-h² for psychological traits systematically converged on &lt;0.05 — has not occurred.

**A2 — GWAS signal is real (not artifact).** Carries the PGS enterprise; genetic nurture estimates; cross-disorder pleiotropy; modern psychiatric genetics. Robustness: MODERATE-HIGH (within-family PGS effects are non-zero for EA, BMI, height — direct signal exists; AM/stratification inflation magnitudes still being quantified). Would flip if within-family PGS effects converged on zero across most traits.

**A3 — `g` is a real dimension of cognitive variation.** Carries Section 5 of lit review; predictive-validity claims; CHC structure; mortality/income predictions. Robustness: HIGH for g as a statistical regularity; MODERATE for g as unitary biological mechanism. Would flip if a broad cognitive battery had first-PC &lt;15% or if interventions reliably moved one ability while lowering others. **2024 wrinkle**: Pietschnig et al. 2024 reported the positive manifold may be *weakening* across recent cohorts — softly pressures A3 in a new way without refuting it.

### 3b. Reframer nodes — the answer is open and reshapes interpretation

These don't break the picture if reversed; they change what the picture *means*. Their magnitudes are being actively quantified in 2024–2026 work. Conflating reframers with foundational cruxes is the most common conceptual error in pop-science treatments of this field.

**G2 / E6 — Passive rGE / genetic nurture.** Reframes the *meaning* of every population-level genetic estimate. Without G2, "genetic transmission" reads as direct biological causation; with G2, ~half is environmentally mediated by genetically-similar parents (Wang 2021 / Isungset 2022). Nivard et al. 2024 (Nat Hum Behav) showed indirect genetic effects extend beyond the nuclear family to dynastic / extended-family processes. The existence is robust; precise magnitude across all traits is still being quantified.

**G6 / E7 — Cross-trait assortative mating.** Reframes the cross-disorder genetic-correlation matrix and the p-factor's interpretation. Border 2022 (*Science*) showed phenotypic cross-mate correlations explain R²=74% of variance in genetic-correlation estimates. Ma, Wang, Border et al. 2024 (LAVA-Knock) is the first method to systematically reduce xAM-induced bias. The share of any specific rg that is artifact vs. genuine pleiotropy is still pending.

### 3c. Logical guardrails — unfalsifiable but load-bearing for interpretation

These cannot be falsified — they are algebraic / definitional truths. They *can* be ignored, which is how most public-discourse misuse of the field happens.

**L1 — Heritability is a population variance ratio, not an individual partition.** Cannot be falsified. Public misreading of "70% heritable IQ" as "70% of any individual's IQ comes from genes" is the failure of L1, not the science.

**L4 — Within-population heritability does not license between-population mean inference (Lewontin firewall).** Cannot be falsified — it is a logical/algebraic point. Can only be ignored. The empirical buttress today is E23 (PGS portability collapse along genetic-distance continuum, Ding 2023): even if you wanted to use within-pop methods to speak to between-pop differences, the methods don't currently work.

### 3d. Decorative material (safe to compress)

Removable from the topology without changing the qualitative picture:

- **E35** (CAH / primate toy preferences) — convergent evidence, not necessary
- **E37** (specific GABAergic critical-period mechanisms) — biologically real, not load-bearing for the variation argument
- **HEXACO Honesty-Humility specifics** — incremental over Big Five
- **Specific Dark-Triad subdimensions** — D-factor synthesis (Moshagen 2018) carries more weight
- **P-FIT brain network specifics** — corroborate g but don't establish it
- **Yehuda Holocaust FKBP5 transgenerational findings** — refuted/non-replicated; kept only as historical anchor for D1 distortion
- **Specific candidate-gene findings (5-HTTLPR depression)** — refuted; kept as historical anchor for the field's methodological turn (M8)

---

## 4. Weakest links

These are the load-bearing pieces with the lowest current confidence. Targeted attack on any one would do the most damage to the integrated picture.

### W1: Generalization from candidate-GxE failure to "all GxE is small" (E5 → broader claim)

**Why fragile**: The candidate-gene collapse is definitive. The extrapolation that polygenic-score × environment interactions are also small is an *inductive leap*, not a result. As of 2025, the picture is **partially holding the null but not strengthening it**. A 2025 systematic review of 56 PGS×E studies for depression found mostly null or small effects. A multivariable PGS×E study of educational achievement (Allegrini et al. 2020) found "no evidence that GxE effects significantly contributed to multivariable prediction." UK Biobank work (2024) on distinct explanations of GxE shows that many apparent GxE signals are confounded by scale, ascertainment, or population structure. The candidate-gene-failure extrapolation is looking less like an inductive leap and more like a substantive empirical pattern — but the literature is still too young for a strong null.

**Pressure test**: Several large preregistered PGS×E studies finding interactions explaining &gt;5% variance would substantially revise this corner of the picture.

### W2: Scarr-Rowe (E25) — has substantially weakened since pass-0

**Why fragile**: The original meta-analytic picture was "replicates in US, fails in W. Europe / Australia" (Tucker-Drob & Bates 2016). Ghirardi et al. 2024 (Netherlands Twin Register, polygenic-index design across 42 PGI×SES interaction tests for educational outcomes) found 39/42 negative, 0 significant positive, 1 marginally significant positive — i.e., the *opposite* sign from Scarr-Rowe in most cases. The picture in 2026 is closer to "the **compensatory hypothesis** (more genetic expression in low-SES because constrained environments suppress non-genetic variance) is the better-supported pattern, at least for educational outcomes." E25's weight has been downgraded from 3 → 2 to reflect this. The narrative "deprivation suppresses heritability" — popular in policy discourse — is now evidence-thin.

### W3: Plomin-vs-Turkheimer interpretation of PGS (O1)

**Why fragile**: Both views are compatible with current data. Determines what PGS *means* — direct biology vs. summary statistic of correlated environments. The field publishes ambiguously across both interpretations. Will likely be settled only by within-family-only PGS that are still well-powered.

### W4: Magnitude of AM-correction across psychiatric cross-disorder rg matrix (O7) — methods now emerging

**Why still fragile but improving**: Border 2022 showed xAM explains R²=74% of variance in genetic-correlation *estimates* but didn't prove all rg's are spurious — some genuine pleiotropy surely exists. As of 2024, the field is moving from flagging the problem to building correction methods. **Ma, Wang, Border et al. 2024** (*American Journal of Human Genetics*) introduced **LAVA-Knock**, a local-genetic-correlation method using knockoff inference to reduce xAM-induced bias; tested across 630 trait pairs in simulation and real GWAS, it substantially reduces but does not eliminate the bias. A 2024 study found AM genetic signatures across SCZ, BD, MDD, alcohol phenotypes, and Tourette syndrome — confirming xAM is not selective. **What's still pending**: how much of the cross-disorder rg matrix and the p-factor genetic signal survives systematic application of AM-correction methods at scale. Likely answer in 2–3 years.

### W5: Gender Equality Paradox mechanism (O3, E14) — pattern strengthened, mechanism still contested

**Empirical pattern: more robust as of 2025**. Herlitz et al. 2025 systematic review (54 articles, 27 meta-analyses, *Perspectives on Psychological Science*) found the paradox replicates across personality, verbal abilities, episodic memory, and negative emotions. Balducci et al. 2024 extended it to within-individual academic strengths cross-temporally. The "this won't replicate" objection has weakened.

**Mechanism: still contested**. Three live candidates: (a) innate-expression release in resource-rich environments, (b) reference-group / self-anchoring artifacts in self-report (people compare to their gender peers, not to humans-in-general), (c) wealth/freedom confounds with gender equality. Behavioral / incentivized-measure replications (Falk & Hermle 2018 for economic preferences) cover only part of the domain. The decisive test — non-self-report behavioral replication across personality and interests — is still incomplete. Each candidate mechanism implies different normative conclusions, which is part of why this remains contested rather than resolved.

### W6: Flynn-reversal cause (O2) — and a new wrinkle on the positive manifold

**Why fragile**: The pattern is environmentally driven (within-sibship, Bratsberg & Rogeberg 2018), so "dysgenic" explanations are out. But no mechanism (screen time, education quality, attention, nutrition, lead, microplastics) has been pinned down with within-cohort empirical work. Pietschnig et al. 2024 (Vienna 2005–2018 cohort) added a structural twist: the *positive manifold itself* may be weakening across cohorts — meaning the recent rise/fall is **not uniformly g-loaded**. If confirmed broadly, this softly pressures A3 (g exists as a stable dimension) — not refuting it, but suggesting its strength may be cohort-dependent. Still not load-bearing for the integrated picture, but interacts with A3 in a new way.

### W7: A6 (dimensional vs. taxonic) at psychiatric extremes

**Why fragile**: Most psychopathology is dimensional (taxometric evidence is robust), but for severe early-onset autism with intellectual disability, rare large-effect variants (CHD8, SCN2A, SYNGAP1) drive a partly taxonic picture. The "all dimensional" framing oversells continuity at the severe tail.

---

## 5. Variant views

The same graph, read four ways.

### Variant A: Vulnerability map — *where does this break?*

The vulnerability map is the union of the three foundational cruxes (§3a), two reframer nodes (§3b), two logical guardrails (§3c), and seven weakest links (§4). Together they describe the smallest set of pressure points whose movement would force restructuring of the integrated picture:

- Falsify A1: SNP-h² systematically &lt;0.05 → twin-method discredited → Section 1 collapses
- Falsify A2: within-family PGS → 0 → modern psychiatric genetics collapses
- Falsify A3: positive manifold dissolves → Section 5 collapses
- Falsify G2: within-family PGS = population PGS → genetic nurture is null → Plomin direct-causal view wins (O1 resolves)
- Falsify G6 fully: AM correction barely changes rg matrix → cross-disorder pleiotropy is real
- Violate L4: cannot be falsified, only ignored — but its violation in public discourse is the largest single source of public confusion

If exactly one of these were to flip, the rebuild would be: A1→ rebuild Section 1 only; A2→ rebuild Sections 1, 3, 7 (~40% of lit review); A3→ rebuild Section 5 (~25%); G2/G6→ keep numbers, rewrite causal interpretation throughout.

### Variant B: Flow map — *how does causation propagate?*

Causation in this system runs in two directions, both important.

**Forward developmental flow** (genome → outcomes):

```
Genome (polygenic + few rare large-effect)
    │
    ├──> Temperament (infant biological reactivity: Surgency / NA / EC)
    │       │
    │       ├──> Active rGE / niche-picking ─────────┐
    │       │                                        │
    │       └──> Evocative rGE (eliciting responses) ┤
    │                                                │
    └──> Direct expression in brain development ─────┤
                                                     ▼
                                                Personality (adult)
                                                     │
                                                     ├──> Attainment
                                                     ├──> Relationships
                                                     ├──> Health behaviors
                                                     └──> Mortality
```

**Indirect / dynastic flow** (parents' genome → offspring environment → offspring outcome):

```
Parents' genome
    │
    ├──> Parents' phenotype (income, vocabulary, parenting style, neighborhood choice)
    │       │
    │       └──> Offspring's rearing environment ────────┐
    │                                                    ▼
    │                                              Offspring outcomes
    │                                                    ▲
    └──> Transmitted alleles ───────────────────────────┘
```

The genetic-nurture finding (E6) says these two pathways have roughly equal magnitude for educational attainment. They are partially separable only via within-family designs (M4) or non-transmitted-allele PGS.

**Cross-generational drift via assortative mating:**

```
Mating choice (correlated on phenotype) 
    │
    └──> LD induction among causal variants (G5)
              │
              ├──> Inflated additive genetic variance
              ├──> Inflated h²
              ├──> Inflated cross-trait genetic correlations (G6)
              └──> Inflated PGS prediction accuracy
```

### Variant C: Minimal claim set — *smallest set supporting the conclusion*

The smallest collection of claims that yields the integrated picture (S1) is **eight nodes**:

1. **E1** — Mean trait h² ≈ 0.49 (heritability is real and substantial)
2. **E4** — Polygenic architecture (no master genes)
3. **E6** — Within-family PGS ≈ ½ population PGS (genetic nurture is real)
4. **E7** — Cross-trait AM is a major source of inflated genetic correlations
5. **E8** — A small set of large environmental insults have causal effects (lead, iodine, alcohol, deprivation, schooling)
6. **L4** — Within-pop ≠ between-pop (Lewontin)
7. **G7** — Stochastic developmental noise is the dominant source of non-shared environment
8. **A6** — Most psychological variation is dimensional, not taxonic

These eight together generate the qualitative integrated picture without requiring detailed effect-size tables, cross-cultural caveats, or specific candidate-gene history. The remaining ~50 nodes refine and corroborate but do not change the shape.

### Variant D: Politicization map — *where does motivated reasoning concentrate?*

This is the variant most relevant to the topic framing ("a minefield of motivated reasoning on all sides").

**Distortion-to-target matrix:**

| Distortion | Targets | Move | Counter-evidence |
|---|---|---|---|
| **D1** Blank-slate | A1, E1, E10–E14 | "Twin studies are flawed; differences are socialization" | SNP-h² (bypasses EEA), MZ-reared-apart, Su 2009 (d=0.93), CAH/primate convergence |
| **D2** Hereditarian | L4, E22, E23, O4 | "Group differences are genetic" | PGS portability collapse (E23); Lewontin (L4); cross-ancestry GWAS at scale don't exist |
| **D3** Gender-similarities | E10, E11, E14 | "All differences are tiny (cite math d=0.05)" | Multivariate D=2.71 (E10); people-things d=0.93 (E11); GEP (E14) |
| **D4** Pop-evpsych | A6, L7, E10–E14 | "Men are X, women are Y" (categorical from dimensional) | A6 (dimensional); L7 (effect-size context) |

**Why all four distortions can target the same evidence base**: the evidence base contains *both* large differences (people-things d=0.93) *and* trivial ones (math d=0.05) *and* strong heritability (h²=0.49) *and* large environmental insults (lead, schooling) *and* logical guardrails against between-group inference (L4). Any single-direction narrative requires selective citation. The integrated picture (S1) requires holding all of it at once.

**Operational implication for the formalization stage**: any model that *only* parameterizes the variance components without parameterizing the *interpretation* of those components will be silently captured by whichever distortion the reader is most prone to. The formal model needs to make L4, G2, G6, and the dimensional/taxonic distinction (A6) structurally visible, not just numerically present.

---

## 6. Topology → formalization handoff

What the next stage (model formalization) should pick up.

### Ready for equations

1. **Variance decomposition** — fully specifiable now:

   V(P) = V(A_direct) + V(A_indirect) + V(A_AM-LD) + V(C_residual) + V(E_measured) + V(E_stochastic) + 2·Cov(G,E) + V(GxE)

   With each V parameterized by trait, age, population (US vs. Europe for E25), and method (twin / SNP / within-family). Cov(G,E) captures rGE; V(A_AM-LD) captures G5/G6; V(A_indirect) captures G2.

2. **Method gradient** (S2): twin h² ≥ SNP h² ≥ within-family h², with the gaps decomposable into AM, rGE, and rare-variant contributions. Parameterize as a function of estimation method.

3. **Wilson-effect curve**: h²(age) = a + b·log(age) or similar saturation form, with the slope driven by G1 (active rGE). Calibratable from Bouchard 2013 and Briley & Tucker-Drob 2013.

4. **Multivariate sex-difference algebra**: D² = (μ₁ - μ₂)ᵀ Σ⁻¹ (μ₁ - μ₂), with a worked example showing how D = 2.71 follows from moderate univariate ds and a positive-correlation covariance structure.

5. **PGS-portability decay function**: prediction accuracy as a continuous function of genetic distance from training population (Ding et al. 2023, *Nature*: r = −0.95 between genetic distance and accuracy across 84 traits).

### Still at observation stage (formalization premature)

- O1 — Plomin/Turkheimer interpretation of PGS: not yet a formal disagreement, just a verbal one
- O3 — GEP mechanism: the algebra of "innate expression release" is not yet specified
- O6 — what non-shared environment *is*: no candidate decomposition
- O7 — share of cross-disorder rg that survives AM correction: empirical question pending — but methods are now emerging (LAVA-Knock); answer likely in 2–3 years, at which point this moves to "ready for equations"

### Connection to adjacent topics in the LLM-iterate pipeline

This topology is the natural input to **Parent-to-Child Transmission** (planned topic). The genetic-nurture finding (G2/E6) and the dynastic-extension finding (Nivard 2024) are the empirical answers to "how much does parenting matter beyond genes" that the parent-child topic will need to build on. When that topic spins up, the variance decomposition equation here should be its starting point.

Less directly: the **Evolution-Modernity Mismatch** topic will lean on the GEP (E14/O3) and Flynn-reversal (O2) findings as evidence of environment-driven shifts in expressed psychological variation. **Bedrock Generating Functions** can read the variance decomposition itself as one such bedrock function.

---

## 7. Next moves — three options for Stage 3

The user picks one of these as the primary formalization target. Each leaves the others viable as later modules but shapes Stage 4 (data) differently.

### Option A — Variance decomposition + method gradient (most central)

Build the central equation `V(P) = V(A_direct) + V(A_indirect) + V(A_AM-LD) + V(C) + V(E_meas) + V(E_stoch) + 2·Cov(G,E) + V(GxE)` parameterized by trait, age, population, and estimation method. Build a tool that takes a published h² estimate (twin, SNP, or within-family) and outputs a method-corrected estimate with explicit AM/rGE adjustment.

**Pros**: most central to the topic; directly answers "what generates psychological variation"; feeds Stage 4 cleanly (every term has published estimates somewhere). **Cons**: many parameters; risk of producing a calculator nobody uses without strong UI judgment. **Stage 4 implication**: pull h² estimates from PGC, SSGAC, GIANT consortia; calibrate the method-gradient term per trait class.

### Option B — Multivariate sex-difference algebra (most pedagogically clean)

Formalize how moderate univariate Cohen's ds combine into a large multivariate Mahalanobis D, with a worked Big-Five example showing how D ≈ 2.71 emerges from |d| ≈ 0.4 ds and a positive-correlation Σ. Build a dashboard letting the user dial univariate ds and the correlation matrix to see D move.

**Pros**: tightly scoped; resolves the single biggest framing trap in the GEP debate (univariate vs. multivariate framings of the same data); high pedagogical leverage. **Cons**: narrower than A; doesn't engage the heritability core. **Stage 4 implication**: pull effect-size matrices from Del Giudice 2012 and Schmitt 2008 cross-cultural data; replicate D under different correlation structures.

### Option C — PGS-portability calibration (most practically useful)

Turn Ding et al. 2023's continuous decay finding into a usable accuracy estimator: enter an individual's genetic distance from the PGS training population and get an accuracy-decay multiplier. Apply across the major trait PGSs (EA, SCZ, BMI, etc.).

**Pros**: directly addresses a real-world bias; smallest scope; ships fastest; useful even outside this project's domain. **Cons**: less central to the heritability question; might fit better as a tool than a topic-stage. **Stage 4 implication**: pull cross-ancestry GWAS validation data from the All of Us / GenomeAsia / H3Africa consortia.

**My recommendation**: A as primary (most central to the topic's stated purpose), with B as a stretch module if scope allows. C is high-value but might better live as a standalone tool promoted to `/models` later.

---

## 8. Objections to this topology (adversarial + steelman)

Four ways a careful reader could push back. The strongest version of each, then my response.

### Objection 1 — Discrete typed edges falsify a continuous, magnitude-weighted, context-conditional system

Heritability is not "supported by" a twin study in the same binary way that a logical implication holds. The system is a tightly coupled developmental process; flattening it into nodes-and-arrows with discrete edge types loses information about magnitude, conditional dependence, and gradient relationships.

**Response**: Acknowledged, and intentional. The topology is the qualitative skeleton; edge weights and conditional dependencies are the job of Stage 3 (formalization), where each edge will be turned into a parameterized function. The graph's value is not that it stands in for the full system but that it makes the structure visible cheaply enough that the formalization knows where to put the parameters.

### Objection 2 — The crux/decorative split is editorial, not empirical

There is no algorithm that picks crux nodes; the choice depends on which failure modes you are worried about. A 1990s topology of this field would have crowned candidate-gene findings as cruxes. Naming A2 (GWAS signal real) a crux today is a judgment call about the field's current methodological commitments — not an objective feature of the science.

**Response**: Correct. Cruxes are time-stamped. This topology is a 2026 snapshot. If the field shifts (post-AM-correction era, post-within-family-PGS-at-scale era) the crux set will shift — that is what the refinement passes are for. Use this as a current map, not an immutable structural claim.

### Objection 3 — Calling L4 a "logical firewall" overstates the case

Lewontin's 1970 argument has been challenged. Edwards (2003) "Lewontin's Fallacy" showed that Lewontin's specific quantitative point — that ~85% of human genetic variance is within rather than between populations — does not preclude reliable population-classification from genetic markers. Modern population genetics treats between-population genetic inference as more nuanced than the firewall framing suggests.

**Response**: The Edwards critique is real, but it addresses a different claim. Edwards refuted "you cannot reliably classify individuals into populations from genetic data." The L4 firewall as I formulate it says "within-population heritability provides no information about between-population *mean differences* without strong auxiliary assumptions about shared causal architecture and equal environments." Those are different propositions. PGS portability collapse (E23) is the contemporary empirical evidence that the auxiliary assumptions are not currently being met for psychological traits. The firewall framing survives the Edwards critique; its strength rests on the empirical PGS-portability finding, not on the original Lewontin variance argument alone.

### Objection 4 — The Politicization variant is meta-commentary, not topology

The D nodes and `attacks` edges describe how people misuse the evidence base. That is epistemics or sociology of science, not structural topology of the field. A pure topology should omit them.

**Response**: Fair, and the inclusion is non-orthodox. It is justified here only by the topic framing — the user's prompt explicitly described the field as "a minefield of motivated reasoning … where the actual generating functions are obscured by politics." A topology of just the science would omit the D nodes; a topology that helps a reader navigate the field as it is actually encountered should include them. The D nodes will not be carried into Stage 3 formalization — they exist for navigation, not for downstream computation.

---

## 9. Glossary

For readers approaching this from outside the field. Terms appear throughout the lit review and topology; this is the lookup table.

| Term | Meaning |
|---|---|
| **h²** | Heritability — fraction of trait variance in a population attributable to genetic variation. A population statistic, not an individual one. |
| **SNP** | Single-nucleotide polymorphism — a single-base difference at a position in the genome where multiple variants exist in the population. |
| **GWAS** | Genome-wide association study — scans hundreds of thousands of SNPs against a measured trait, looking for statistical association. |
| **PGS** | Polygenic score — a per-individual sum of trait-associated SNPs weighted by their GWAS effect sizes. Used as a predictor. |
| **LD** | Linkage disequilibrium — non-random association between alleles at nearby loci, typically because they are inherited together. |
| **AM** | Assortative mating — partners resemble each other on a trait above chance. **xAM** = cross-trait AM (e.g., taller-than-average partners with more-educated-than-average). |
| **rGE** | Gene-environment correlation. **Passive** (parents transmit genes + correlated environment), **evocative** (heritable traits elicit responses), **active** (people select environments matching propensities). |
| **GxE** | Gene-environment interaction — the same genotype produces different phenotypes in different environments. |
| **EEA** | Equal environments assumption — the twin-method assumption that MZ and DZ twins are treated similarly enough that any extra MZ phenotypic resemblance reflects genetics, not differential treatment. |
| **MZ / DZ** | Monozygotic (identical, ~100% shared DNA) / dizygotic (fraternal, ~50% shared DNA) twins. |
| **rg** | Genetic correlation between two traits — how much the same genetic variants influence both. |
| **WGS** | Whole-genome sequencing — capturing every base in the genome, including rare variants GWAS misses. |
| **g-factor** | General factor of cognitive ability — the latent dimension behind the positive manifold (every cognitive test correlates positively with every other). |
| **p-factor** | Proposed general factor of psychopathology — analogous to g, derived from cross-syndrome correlations. |
| **CHC / HiTOP** | Cattell-Horn-Carroll cognitive-ability hierarchy / Hierarchical Taxonomy of Psychopathology (a dimensional alternative to DSM). |
| **d (Cohen's d)** | Standardized mean difference between two groups, in standard-deviation units. Effect-size labels (small / medium / large) are scale-dependent — see L7 in the node catalog. |
| **Mahalanobis D** | Multivariate generalization of Cohen's d — distance between two group means in the geometry of the trait space, accounting for correlation between traits. |
| **Within-family design** | Comparing siblings or MZ-discordant twins or parent-offspring trios within the same family — controls for between-family confounds (population stratification, AM, passive rGE). |
| **Genetic nurture** | Effect of *parents'* genotype on offspring outcomes via the environment the parents create — including alleles the parent did not transmit. |

---

## 10. Stage_outputs convention reference

Raw working drafts from each LLM-iterate stage live at:

```
stage_outputs/<topic>/<stage>.md
```

Where `<topic>` is kebab-case (e.g., `human-psych-variation`) and `<stage>` is one of: `lit-review`, `topology`, `model`, `data`, `build`, `writeup`. Polished versions move into `src/content/ai_research/<topic>/<stage>.mdx` with proper frontmatter (title, description, date, status, refinementPass, refinementLog) once ready to publish on the site.

The interactive D3 graph for this topology lives at `src/components/research/PsychVariationGraph.tsx` and is mounted in `src/content/ai_research/human-psych-variation/topology.mdx` via `client:load`.

---

## Navigating an AI World
*topic: navigating-ai-world · stage: overview · pass 0 · in-progress*

How the AI transition restructures work, personal life, and relationships — not surface-level advice but a structural analysis of what changes and why. The output should be useful for someone making real decisions about career, education, relationships, and life architecture as AI capability accelerates.

The topic running through the LLM Iterate pipeline. The point is not a take but a structural map: which mechanisms are doing the work underneath the discourse on labor, on relationships, and on meaning, and how those mechanisms interact.

Stage 1 (lit review) assembles the three domains: labor economics through 2025-2026, the empirical companion-app and therapy-bot literature, and the philosophical/psychological literatures on identity and meaning that the AI case forces us to repurpose.

Stage 2 (topology) is the dependency graph — what depends on what, which nodes are foundational cruxes vs. reframer mechanisms vs. logical guardrails, where the weakest links are, and where individual decision-leverage actually concentrates.

Stages 3–6 will formalize a usable model, run it against available data, ship a tool, and write the long-form synthesis for an educated lay reader.