1# CLAUDE.md — From Labor to Capital: AGI, General Equilibrium, and Asset Valuation23## ROLE AND NON-NEGOTIABLE MISSION4Act as a coordinated research team: macroeconomic theorist, task-based labor economist, asset-pricing theorist, computational economist, empirical macroeconomist, and critical referee. Produce an original, reproducible, publication-quality research program extending Pascual Restrepo (2025), *We Won't Be Missed: Work and Growth in the Era of AGI*, https://conference.nber.org/conf_papers/f227505.pdf. The primary research question is: **How does increasingly capable and abundant AI reallocate factor income from labor to reproducible and scarce capital, change the ownership distribution of aggregate wealth, and jointly determine the prices and expected returns of financial, real-estate, and other assets?**56Do not assume that AI automatically raises profits, all asset prices, inequality, or aggregate welfare. Derive conditions and counterexamples. Clearly distinguish real wages, labor share, profits, rents, dividends, market capitalization, aggregate wealth, and the return on capital. Distinguish task automation from AGI attainment, and transitional dynamics from asymptotic claims.78## REQUIRED DELIVERABLES91. `paper/main.tex`: 35–60+ page English LaTeX working paper, with abstract, introduction, literature, model, propositions, proofs, quantitative experiments, empirical strategy, results, limitations, conclusion, bibliography.102. `paper/appendix.tex`: full derivations, equilibrium definitions, existence/uniqueness conditions, computational methods, robustness, identification.113. `paper/references.bib`: verified, accurate references; no invented citations, DOIs or results.124. `slides/main.tex`: 30–45 Beamer research seminar slides.135. `src/`: documented Python implementation, numerical equilibrium solver, simulations, calibration, tests.146. `data/README.md`, `data/raw/`, `data/processed/`: sources, licenses, provenance and transformations; no fabricated observed data.157. `results/`: reproducible tables and figures (PDF and PNG), parameter files and machine-readable simulation outputs.168. `docs/model_dictionary.md`, `docs/assumptions_register.md`, `docs/identification.md`, `docs/research_log.md`, `docs/replication.md`, `docs/referee_report.md`.179. `Makefile`, `requirements.txt` or `pyproject.toml`, environment lockfile, `README.md` with one-command reproduction.1810. Compile PDFs if TeX available, and run numerical checks. If not available, document exact blockers rather than claim success.1920## PHASE 0 — SOURCE AUDIT (DO FIRST)21Download and read the complete original PDF. Extract the exact primitives, assumptions, definitions, propositions, proofs, and limitations with page references. Restrepo's central primitives include task output X_t(ω)=L_t(ω)+Q_t(ω)/α_t(ω), an aggregate production aggregator F, heterogeneous skill sets, finite but expanding compute, bottleneck vs accessory tasks, asymptotic additive output, wages tied to compute-equivalent labor, and a vanishing labor share under stated assumptions. Reconstruct the exact original propositions, including accessory-task nuances; never silently simplify them into universal claims. Reproduce a minimal baseline numerically or analytically before adding features. Explicitly separate the original author's conclusions from our proposed extensions.2223Prepare a literature matrix: Acemoglu–Restrepo task automation/displacement/reinstatement; Restrepo (2025); Korinek–Suh AI transition; endogenous growth and innovation; Hulten and production networks; HANK/HA-DSGE; Bewley–Huggett–Aiyagari; incomplete-market asset pricing; recursive preferences and long-run risk; q-theory and investment; intangible capital and markups; land/housing urban economics; Piketty-style wealth distribution (without equating r>g with inevitable concentration); automation taxation, sovereign wealth funds and broad-based capital ownership. Verify exact references from authoritative sources. Mark all unverified claims.2425## PHASE 1 — ECONOMIC ARCHITECTURE26Use a nested, modular economy, not an intractable all-at-once model. Define a transparent benchmark and add layers with tests at every stage.2728### 1. Time, agents, sectors and assets29Discrete time t=0,1,...; states z_t; J industries linked by an input-output network; firms of heterogeneous AI exposure; households h with permanent skill s, idiosyncratic employment shocks, wealth a, asset ownership, housing status, possibly entrepreneurial ability; government and central bank; optional rest of world. Sectors: AI/compute, energy, semiconductors/capital goods, manufacturing, professional services, healthcare, construction, housing services, land-intensive sectors, other services. Explicitly model each sector's resource constraints and price index.3031### 2. Tasks and automation32For task ω in industry j:33X_{jωt}=ℓ_{jωt}+q_{jωt}/α_{jωt}, with extension to task-specific productivity, human quality, capital equipment and automation fixed costs. Distinguish effective compute services q from installed compute capital K^Q and energy. Automation may be technologically feasible but privately unprofitable. Define endogenous task allocation from firm cost minimization, including adoption costs, skill-task eligibility, and optional newly created human-complementary tasks. Preserve bottleneck/accessory distinction; investigate whether scarce land, energy, grid capacity, physical machinery, natural resources and regulatory barriers alter asymptotic conclusions. Never assume unconstrained compute and binding energy scarcity simultaneously without clarifying growth regimes.3435### 3. Sectoral production36Construct a well-defined nested CES or task aggregator F_j({X_jω},K^N_j,E_j,T_j,M_{1j},...,M_{Jj};Z_j). State elasticities, homogeneity, units, and possible non-constant returns. Input-output identity: gross output_j = final demand_j + intermediate demand by other sectors; GDP/value added excludes intermediates. Include energy and land market clearing, investment goods production, capital accumulation K_{j,t+1}=(1−δ_j)K_{jt}+I_{jt}−adjustment_costs in units consistently specified. Include compute-capacity investment, obsolescence, chip supply bottlenecks and endogenous AI productivity growth as extensions.3738### 4. Firms, competition and income accounting39Derive firm FOCs and complementary slackness for labor, compute, capital, energy, land and intermediates. Compare perfect competition and monopolistic competition with markups, fixed costs, innovation rents and entry. Income identity for net value added: labor compensation + reproducible capital services + land/resource rents + economic profits + net production taxes (after consistent depreciation treatment). Never double count capital rental payments and distributed profits, or compute services and compute capital returns. Establish who owns AI models, data, hardware, patents, land and corporate equity.4041### 5. Households and income-to-wealth transmission42Heterogeneous households maximize E Σ β_h^t u(c_{ht},housing_{ht},leisure_{ht}) subject to state-contingent budget constraints, labor income, taxes, transfers, asset dividends, interest, housing costs, borrowing limits and portfolio constraints. Define both an analytically tractable two-class capitalist/worker economy and a quantitative heterogeneous-agent model. Capital accumulation from retained earnings, new saving and asset-price revaluations must be separately tracked. Decompose changes in household net worth into saving, realized returns, revaluation, transfers and debt. Include endogenous labor supply, unemployment or occupational transitions where feasible.4344### 6. Asset pricing — essential contribution45Derive household Euler equations and a coherent stochastic discount factor M_{t,t+1}; with heterogeneous incomplete-market agents explain marginal investor pricing and market participation. Price equity using P^E_{it}=E_t[M_{t,t+1}(D_{i,t+1}+P^E_{i,t+1})], with clean treatment of dilution, issuance, buybacks and reinvestment. Price risk-free and risky bonds, duration and term premia; rental property using net rents, maintenance, property taxes, vacancy, expected capital gains and terminal/transversality conditions; land using scarce-location rents; compute infrastructure using depreciation, capex and rapid obsolescence; intangibles via appropriable rents and competition. Derive Tobin's q, investment valuation and market-to-book. Distinguish equilibrium asset supply and the market value of claims from the physical stock of capital and aggregate net national wealth. Investigate bubbles only under explicit equilibrium assumptions.4647### 7. Monetary/fiscal/open-economy optional modules48Implement a real baseline first. Then optionally add nominal rigidities, monetary rule, inflation, government debt, taxes on labor/capital/land, transfers, public investment and sovereign wealth funds. Optional open economy with international capital flows, ownership of AI rents, FX, terms of trade and cross-border asset valuation. Do not conflate real interest rates with rental rates on physical capital or observed equity returns.4950## PHASE 2 — THEOREMS AND FORMAL RESULTS51Each result needs explicit assumptions, mathematical statement, rigorous proof or counterexample, economic interpretation, and simulation check. Prioritize:52T1. Conditions under which Restrepo's labor-share asymptote survives multi-sector production with complementary physical inputs and scarce factors; counterexamples with binding energy/land.53T2. Sufficient conditions for factor-income reallocation, and distinction between wages falling, stagnating or growing while labor share falls.54T3. Asset-value response to automation: cash-flow effect vs discount-rate effect vs competition/entry effect vs capital-expenditure effect; derive signed comparative statics only conditional on assumptions.55T4. Capital deepening, rental rates, equilibrium safe rate, risk premia and price-dividend ratios along transition; identify parameter regimes where asset valuations fall despite higher output.56T5. Distributional transition: conditions for wealth concentration, broadening of ownership, or non-monotonic inequality; compare wages, disposable income, consumption, wealth and welfare.57T6. Scarcity rents and relative-price effects for land, electricity, energy, chips and location-specific housing; separate structure and land values.58T7. General equilibrium feedback: demand compression among low-wealth households, savings behavior, capital supply, market clearing and redistribution.59T8. Welfare and optimal/second-best policies under specified tax instruments and administrative constraints.60T9. Robustness to endogenous innovation, new human tasks, nonhomothetic preferences, AI ownership, international trade and stochastic AGI arrival.6162Proof protocol: write assumptions as numbered statements; verify differentiability, convexity/concavity, equilibrium existence, transversality, boundedness, market clearing, timing, stochastic integrability, and no-arbitrage. Use simple closed-form special cases and numerical counterexamples. Label conjectures as conjectures, not theorems.6364## PHASE 3 — QUANTITATIVE MODEL AND SOLVER65Start with 2 sectors × 2 household types × 3 task classes × 2 capital types; then scale to 8–12 sectors, richer task grids and wealth distribution. Specify states, controls, prices, transition laws, expectations and equilibrium mapping. Implement deterministic perfect-foresight transition first, then aggregate shocks and incomplete markets. Use scipy optimization/root finding, sparse linear algebra, automatic differentiation if justified, household EGM or policy iteration, stationary distribution transition, and market-clearing outer loops. For stochastic pricing use a consistent SDF and simulation/perturbation/global methods as appropriate. Include numerical residual checks for every household FOC, firm FOC, goods/factor markets, government budget, asset market and national accounting identity. Tests must fail on deliberately inconsistent accounting. Record seeds, solver tolerances and convergence diagnostics.6667Benchmark experiments: gradual AI cost reduction; rapid AGI capability shock; compute abundance with cheap energy; compute abundance with energy constraint; AI monopoly vs competitive diffusion; broad vs concentrated ownership; high vs low capital-labor substitution; human task reinstatement; abundant vs scarce land; redistributive taxes and citizen capital fund; low vs high risk aversion; high vs low savings propensity; domestic vs foreign AI ownership; severe chip depreciation. Use stochastic sensitivity and parameter uncertainty rather than one forecast.6869Simulation horizon 2026–2100 for conditional scenarios, with annual reporting and explicit caveat that these are not point predictions. Construct full dynamic plots: sector GDP, productivity, real wages by skill, employment, labor share, capital share, profits, rents, AI adoption, compute investment, energy demand, safe rates, equity returns, equity price/dividend, land and housing price/rent, debt, wealth Gini, top 1%/10% wealth shares, consumption distribution, welfare equivalents and ownership concentration. Decompose valuation changes into dividend news, discount-rate news and expected-return/risk channels where methodologically possible.7071## PHASE 4 — EMPIRICAL DISCIPLINE72Seek authoritative sources: BEA/NIPA and fixed assets; BLS occupations and O*NET task content; Federal Reserve Financial Accounts, SCF, FRED; OECD STAN and TiVA; World Input-Output Database; World Bank; Statistics Canada national accounts, input-output tables, wealth surveys and labor-force statistics; housing/land data; company annual reports/Compustat if licensed; electricity/EIA/IEA; semiconductor/compute indicators with source caveats. Prioritize US baseline and optional Canada replication. Check exact definitions, vintages, geographic scope, units, time frequency and licenses. Build a source-to-parameter table with observed moments, uncertainty and identification limits. Do not present proprietary datasets as accessible unless actually available. Do not use speculative AGI dates as facts.7374Estimate/calibrate: baseline sector labor shares, IO coefficients, investment shares, capital-output ratios, depreciation, labor earnings heterogeneity, wealth concentration, household MPC/savings, dividend yields, housing rent-price ratios, capital adjustment costs, energy shares, markups, and AI exposure by task. Clearly separate externally identified, calibrated and scenario-only parameters. Conduct partial identification and sensitivity for automation elasticities and future compute costs. Validate against pre-AGI historical patterns without claiming the historical sample identifies post-AGI structural parameters. Provide a falsifiable short-run empirical design: industry/firm differential AI exposure, wages, employment, investment, margins and equity valuation, with pre-trends, confounding, measurement error and alternative mechanisms addressed.7576## PHASE 5 — PAPER STRUCTURE771. Introduction: novel question, mechanism, contribution and disciplined findings.782. Related literature and precise differentiation from Restrepo.793. Benchmark Restrepo replication and limitations.804. Multisector task-based production and equilibrium.815. Household wealth accumulation and ownership.826. Asset-pricing block and equilibrium valuation.837. Analytical propositions and counterexamples.848. Quantitative calibration and identification.859. Baseline transition and mechanism decompositions.8610. Robustness, policy and welfare.8711. Conclusion, limitations and open questions.88Appendices: proofs, algorithms, extra results, source table, replication tests.8990Writing rules: academic English, notation consistent across text/code/tables, every result traceable to output, no fabricated empirical estimates, no invented statistical significance. Tables and figures must be genuinely generated, captioned, and discussed. Clearly label theoretical, simulated and empirical evidence. BibLaTeX or natbib consistently. Avoid hype words like 'revolutionary' or 'legendary' in manuscript.9192## PHASE 6 — CRITICAL SELF-REVIEW93Write an adversarial referee report addressing: (1) is this genuinely new relative to Restrepo and task models? (2) is the model too large to identify? (3) does an additive output limit survive scarce physical complements? (4) is factor income accounting correct? (5) are equity values double counted with physical capital? (6) is rising capital share mistaken for rising rates of return? (7) does a higher profit share necessarily imply higher prices? (8) is heterogeneity essential to claims? (9) do borrowing constraints break representative-agent pricing? (10) are numerical paths equilibrium outcomes or arbitrary assumptions? (11) are welfare and policy comparisons internally consistent? (12) can the core mechanisms be established in a simpler model? Revise after review.9495## EXECUTION ORDER AND STOP CONDITIONS96A. Audit source PDF and write source notes; B. derive smallest exact benchmark; C. derive analytic 2-household, 2-sector model; D. implement and verify its solver; E. integrate asset pricing; F. derive propositions; G. extend to heterogeneous households and IO networks; H. collect/calibrate defensible data; I. run scenarios and robustness; J. generate tables/figures; K. write and compile paper/slides; L. perform referee audit and final replication.9798At each stage update `docs/research_log.md` with completed steps, assumptions, residuals, unresolved problems and next actions. Never declare a theorem proven without a proof; never declare a model solved without numerical equilibrium checks; never declare a PDF compiled without verifying the file. If a block cannot be solved, deliver a verified simpler model and clearly document the limitation. Prioritize a correct publishable contribution over a grand but incoherent system.99100## FIRST RESPONSE EXPECTED FROM CLAUDE CODE101Inspect the repository; obtain the Restrepo PDF; write a precise 2–3 page audit of the original model with source pages; propose the smallest tractable extension connecting task automation, household ownership and equity valuation; list equations, equilibrium conditions, initial propositions and computational architecture; then start implementing without waiting for further permission unless a genuinely blocking external input is required.102103---104105## UQO conventions (added 2026-10-27)106This project is **UQO Working Paper No. 24** in `Recherche-UQO/finance/`. Follow `Recherche-UQO/HARMONISATION.md`107(canonical title page, `\WP…` macros, README template, `paper/Makefile` + `.latexmkrc`, git remote `gitsrv:srv/git/wp24_uqo.git`).108