# Adversarial referee report — UQO WP 24, "From Labor to Capital" (v1.0, 2026-10-27)

Written after the full pipeline ran (26 two-class scenarios attempted, 22 solved to tolerance, 4 not — documented in Appendix C; 3 heterogeneous-agent runs). Each of the twelve questions required by the research protocol is answered with a verdict, the evidence, and the revision made or the limitation acknowledged in the manuscript.

# 1. Is this genuinely new relative to Restrepo and task models?

Verdict: yes on three points, no on the production block itself. The work equation, the skill partition and the compute-equivalent wage are Restrepo's. What is new: (i) compute as a reproducible, priced, depreciating, energy-using capital good rather than an endowment — this is what turns "compute stays scarce" from a premise into a result that depends on elasticities and price floors (Propositions 3–5); (ii) the separation of income shares from asset values (Proposition 6: the wealth-to-income ratio depends on the duration of the claims that capture income, and compute is a short-duration claim); (iii) the household and ownership side (two-class and heterogeneous-agent economies, citizen fund, land vs compute taxes). Moll–Rachel–Restrepo (2022) is the closest antecedent on distribution; Korinek–Suh (2024) on the wage collapse. The regime map (Table 1 of Section 7) is the compact statement of what is new. Revision: the introduction now states explicitly which propositions reproduce the source and which depart from it.

# 2. Is the model too large to identify?

Verdict: partly. The dynamic model has four capital stocks, three elasticities, two floors and three rates of technological change. The base point is pinned exactly by shares (Section 8); the elasticities are literature values; the floors and rates are scenario inputs that nothing identifies. The paper does not claim to estimate them. The honest statement is in docs/identification.md and Section 8.2: the level of wages after 2060 is proportional to the product of the compute-price floor and the terminal compute per worker. The scenario catalogue spans the floors (0.12, 0.01, 0.001, 1e-6) and the terminal requirements (×1, ×5). Revision: Table 4 and the "what the data do not pin" subsection were added; the abstract says "conditional scenarios".

# 3. Does an additive output limit survive scarce physical complements?

Verdict: no, and the paper says so. Proposition 2 shows that with σ_T < 1 output is bounded by the scarce input and the price of compute goes to zero; Restrepo's additive limit is the σ > 1 case. The quantitative baseline (σ_T = 0.6) converges to a balanced path where land captures 57–65 % of income and growth returns to the trend rate g_Z. The additive limit is a knife-edge of the source's premise that compute is the only scarce input.

# 4. Is factor income accounting correct?

Verdict: yes, tested. GDP = Y + housing services; compute services and energy are intermediate; the identity GDP = wages + rentals + rents (+ markup profits in the monopoly scenario) holds to 1e-8 at every date of every scenario (tests/test_static.py::test_income_identity_holds, tab08_diagnostics), and the test that counts compute services as final output fails as it should. Depreciation is gross in all rentals (gross income shares are reported; net shares would lower compute's share by δ_Q p K /GDP).

# 5. Are equity values double counted with physical capital?

Verdict: no. V^k = q^k K^k_{t+1} is the market value of the claim on the stock; aggregate wealth is the sum of claims plus land, never the sum of claims and stocks. The no-arbitrage identity (1+r)W_t = payouts_{t+1} + W_{t+1} is tested for every tax-free scenario (test_saved_scenario_is_an_equilibrium). Replacement-cost ratios q/p are reported rather than historical-cost book ratios.

# 6. Is rising capital share mistaken for rising rates of return?

Verdict: no; the paper separates them. The compute share of income rises while the price of compute A falls and the required return r rises during the boom and returns to r* afterwards (Table 3). Proposition 6 makes the point formally: shares, user costs and valuation ratios are three different objects.

# 7. Does a higher profit share imply higher prices?

Verdict: not in general, and the model's own results show the opposite during the boom. Wealth/GDP falls from 3.9–4.3 in 2026 to its minimum around 2040–2050 while income shifts to capital, because (i) the real rate rises (Proposition 7, γ = 2) and (ii) compute is capitalized at a short duration. Valuation ratios recover only when land — a long-duration claim — captures the income. The Campbell–Shiller decomposition (Table 5) attributes the 2026 valuation change of each claim to the two channels.

# 8. Is heterogeneity essential to the claims?

Verdict: essential for the distributional claims, not for the pricing claims. All asset-pricing propositions use only the marginal investor's Euler equation; the two-class model nests the heterogeneous one. The HA block adds: the path of the wealth Gini and top shares, the fraction of households for whom the transfer matters, and a check that the relaxation-based GE return path is close to the two-class one (Table 7). The top-1 % share is not matched (no return heterogeneity) — stated as a known miss.

# 9. Do borrowing constraints break representative-agent pricing?

Verdict: no, by construction, and the paper says why. All claims are one composite asset with a common return; constrained households do not hold it, and unconstrained households satisfy the Euler equation with equality, so the kernel is theirs. Constraints change who saves and hence the level of r, not the pricing formula. With several risky assets and idiosyncratic risk the answer could differ (Constantinides–Duffie); the paper has no aggregate risk, so there is no risk premium to be broken.

# 10. Are numerical paths equilibrium outcomes or arbitrary assumptions?

Verdict: equilibrium outcomes for the 22 solved scenarios, with two honest caveats. The four unsolved cells (Cobb–Douglas, substitutes, energy-constrained, Moore without end) are reported as unsolved and the text relies on the propositions for them. Every path satisfies the four capital Euler equations, the goods market and the capitalists' Euler equation to 1e-9 (max residual), passes the income identity and the no-arbitrage identity, and is insensitive to the horizon (50 extra years change r by less than 1e-8 and log K by less than 1e-7 over the first 225 years). The caveat is the 2026 "news": the whole path is anticipated from t = 0, which produces an immediate investment boom and a jump of the real rate to 6–7 %. The paper reports this as a prediction of the perfect-foresight structure, not a fact.

# 11. Are welfare and policy comparisons internally consistent?

Verdict: consistent within the two-class model. Consumption-equivalent variations are computed from the same utility function that generates the Euler equation, over the full path, for workers (hand-to-mouth, so their welfare is the wage-plus-transfer path) and capitalists. The balanced-budget assumption and the lump-sum rebate are stated. What is not done: optimal policy; the comparisons are of fixed instruments.

# 12. Can the core mechanisms be established in a simpler model?

Verdict: yes — and they are. Propositions 1–8 are proved in the one-complement limit economy of Section 7 with closed forms; the regime map (fig02) is computed with the static block alone. The quantitative model adds magnitudes, timing and the energy/housing/land split. A reader who stops after Section 7 has the paper's mechanisms; the rest is calibration and conditional quantification.

# Additional points raised by the referee pass and how they were handled

• Accessory tasks were dropped from the dynamic model. With a linear accessory block and an energy constraint the marginal value of compute is bounded below and compute absorbs all energy. The paper keeps accessory tasks in the replication only and says so (Section 3, research log).
• Wage collapse magnitude. Wages per worker fall to a few percent of 2026 levels by 2100 in every scenario in which α → 1 and the compute price falls by a factor of 8 or more. This is the arithmetic of the compute-equivalent wage, not a numerical artifact; the ×5 terminal-requirement scenario and the slow-capability scenario show the dependence. The abstract and conclusion state the result as conditional.
• Perfect foresight. No risk premia; the one-period lottery illustration (Section 10) is labelled as an approximation. A full stochastic-arrival solution is listed as future work.
• Energy as electricity. The energy block is interpreted as electricity; fuels and transport are not modelled.