# Hostile referee report — WP 25 v1.0 (self-review, 2026-10-27)

Written as a skeptical referee for a top general-interest or computational-economics journal would, before submission. Items marked [open] are not fixed in v1.0 and are disclosed in the paper's limitations.

# Summary verdict

Interesting and unusually honest paper: the central "efficiency benchmark" is a negative result that undercuts the obvious policy story, and the authors derive rather than assert the externality. The theory is tight but second-order; the quantitative sections rest on illustrative parameters; the agent laboratory is small. Major revision.

# Major comments

1. Is Proposition 3 just Angeletos–Pavan (2007) / Colombo–Femminis–Pavan (2014) relabelled? The use-of-information efficiency under U_xX + U_XX = 0 is theirs. The acquisition of a correlated signal by a measure-zero agent differs from private-signal acquisition (the "knowing what others know" term). The paper should (and does, in the proof) show the identity explicitly rather than cite; the referee would still like a one-paragraph argument for why the correlation externality nets out (answer: the agent's private valuation of the signal already includes its covariance with the aggregate, and with U_X ≡ 0 the aggregate's second-order effect on others is pure transfer). Addressed in Section 4.4; could be sharpened.
2. Second-order approximation. All welfare statements are to second order around a symmetric steady state. With large adoption swings (D from 0 to 1) and strongly strategic substitutes (alpha = −1.33) the approximation may be poor. [open] No global check in the non-linear economy is provided for the portfolio mapping (the non-linear steady state is used, but the game is LQ).
3. Which decision is delegated is a modelling choice. The paper admits the savings channel is dead and switches to portfolio choice. A referee will ask whether firm input choices (misallocation) would give a different ranking of channels. [open] Flagged as the first extension.
4. Information parameters are made up. sig_theta_R = 0.06, sig_H_R = 0.10, sig_A_R = 0.04 drive the 3.3 % of consumption number and the 7.7 % wage-noise s.d. The paper sweeps them and calls them illustrative; a referee will still object to any headline number. Mitigation: headline numbers are presented as orders of magnitude; the qualitative results (PV decreasing in D, D²/M scaling, zero wedge) are parameter-free.
5. Model B decision-quality returns are reduced-form. The mapping from the Model A information friction to the return term (m_H, sigma_rH, sigma_rA) is not derived. [open]; the paper says so.
6. Concentrated ownership is bolted on. The owner class is outside the household block (a point mass). A referee may ask for an explicit two-type household economy. Defensible as Proposition 6's two-class economy, but the owner class does not consume or save optimally. [open]
7. Model C's market maker is ad hoc. Depth, tick and inventory limits are parameters; the H4 amplification depends on them and on the call-vs-CDA design. The paper is explicit that this is the finding (execution matters), but a referee will want the amplification mapped to Proposition 7's lambda_r h D (1−m)/m quantitatively. [open]: only qualitative correspondence is given.
8. The stress test needed engineering. Dividend persistence 0.95, preloaded leverage, auto-repay off, −10 s.d. shock, thin ladder: five choices to produce a spiral. The paper discloses them in the table notes, but the referee will read this as "amplification can be produced when you set things up for it". The honest framing — the transient model error alone does not cascade; it needs persistence, leverage already in place and thin depth — should be the headline of that subsection. Partly addressed.
9. LLM laboratory is small (8 agents × 24 periods × 3 seeds in the quick design) and uses 4-bit 7B/3B models. Results about model identity vs prompt are suggestive. The claim "LLM agents trade more under a shared signal" must be verified against the final cells.csv before submission. [check before submission]
10. Effective M. Spiro 2026-type cross-vendor correlation (~0.77) implies the effective M is near 1 even with many vendors; the paper treats M as illustrative but could calibrate an effective M from reported error correlations. [open]
11. Calibration deviations (beta 0.96/0.95 vs ledger 0.99; alpha 0.33 vs 0.354; rho_e/sigma_e vs Flodén–Lindé) are documented but not justified beyond convention. Minor.
12. Code-paper consistency. All numbers are macros generated from results; figures are generated by scripts; the git revision is printed. The referee could not find invented numbers. Good.

# Minor comments

• Notation: alpha (strategic) vs alpha_K (capital share) fixed in v1.0; nu is used for both Frisch and the earnings innovation in docs/model_specification.md (paper uses eta for the innovation... check).
• The literature section is long (2,100 words); fine for a working paper.
• The architecture figure is dense.
• The bibliography (144 entries) is verified; the Holmström–Milgrom DOI contains an underscore (aer.bst does not print DOIs, so no compile issue).

# Would I recommend publication after revision?

Yes, conditional on (2) a global check of the LQ welfare statements, (7) a quantitative mapping of the lab amplification to the multiplier, and (9) verified LLM numbers. The efficiency benchmark and the D²/M loading are the publishable core; the compute-market crowding-out is a clean negative result; the distributional results are expected but cleanly decomposed.