# Clean top-five referee report: *The Task Frontier* Date: 2026-07-15 Version read: 45-page PDF built from `paper_v3/task_frontier_v3.tex` Procedure: clean read of the current PDF and proof appendix without consulting the v3.7 referee memo before reaching the verdict. ## Recommendation **Revise and resubmit (major revision).** This is now a serious theory paper with a recognizable central mechanism. The paper derives a two-dimensional task frontier from serial production with retries, turns a rare-event certification problem into a direction-specific imitation lag, and uses that lag to price investment direction. The optimal-rotation and repeated-race results are the strongest parts. The current version clears the threshold for an R&R rather than a reject because the formal claims are scoped, the proofs are coherent, and the empirical section no longer treats suggestive aggregates as calibration. I would not recommend acceptance yet. The remaining issues concern economic closure and interpretation, not a hidden algebraic error. ## Major comments ### 1. Clarify the equilibrium composition The introduction presents one chain of equilibrium objects, but the formal paper is a sequence of conditional modules. Task-market pricing determines gap profits; certification and entry are solved conditional on post-certificate values; the leader program holds the episode's lags and research capacity fixed; and the renewal race holds the outer task-frontier state fixed for a cycle. There is no single global Markov-perfect equilibrium jointly determining entry, verification, frontier effort, direction, and evolving task values. That is not fatal, and a modular partial-equilibrium paper can be valuable. The paper should state the composition and its limits once, early and unambiguously. In particular, it should distinguish a downstream use of a previously solved conditional object from a jointly closed equilibrium. The current scope notes are dispersed and easy to miss. ### 2. The horizon-copying assumption is stronger than the results require The phrase “one demonstration reveals a plan” is an intentionally sharp benchmark, but it invites an empirical objection that is unnecessary for the theory. The rotation, racing, and wedge results need a shorter plan-access lag than the tail-certification lag; they do not require literal one-shot copying. State this robustness explicitly. Otherwise readers may treat a modeling normalization as the paper's load-bearing empirical claim. ### 3. Make the welfare comparison a benefit-side wedge unless costs are common The exact accounting in Theorem 7 is correct: private marginal benefit equals gross social marginal benefit net of the post-absorption stream. But the claim that the market “favors” one direction relative to the planner translates from marginal-benefit ratios into choices only when the planner and laboratory face the same research technology and relevant direction-specific marginal costs. With research-cost wedges, the appropriation statistic is only one component of the direction distortion. Add the common-cost condition and label `D` a benefit-side appropriation wedge. This will also make the policy discussion cleaner, because liability and public verification then alter a well-defined component of incentives rather than silently solving the full planner's problem. ### 4. The top-five contribution rests on the combination, so keep standard pieces in supporting roles The representation result, Bertrand surplus-gap identity, two-point testing bound, and convex verification problem are individually familiar. The paper is persuasive when they are treated as steps leading to the rotation and strategic-direction results, and less persuasive when each looks like a separate headline. The revised introduction already moves in the right direction. The theorem hierarchy is now reasonable: the standard price identity is a lemma, the simulation calculation is a proposition, and the main dynamic claims retain theorem status. Continue to present Theorems 5--7 as the economic payoff of the earlier construction. ### 5. The empirical discipline is appropriate, but the missing lag remains the central risk The manuscript correctly refuses to map an aggregate open/closed capability gap into either model lag. That honesty strengthens the paper. It also means the central appropriability object is presently unmeasured. For a theory paper this is acceptable at R&R, provided the conclusion does not imply validation. The task-level reliability evidence should remain descriptive; a sigmoid in aggregate environment-learning time would not validate the paper's execution kernel. ## Technical assessment - I found no fatal error in the task threshold, chain representation, certification lower bound, verification comparative statics, lag condition, annuity accounting, rotation ordering, renewal equilibrium classification, or welfare accounting. - The adaptive two-policy selection claim should be stated as an order result with an adaptive transcript argument; the current revision does this. - The representation theorem should retain its finite/compact scope and should not claim equivalence for arbitrary smooth task-specific success kernels. The current revision does this. - The Bertrand result needs the no-purchase option in the runner-up set. The current revision now makes that condition explicit. - The smooth rising-lag term and the post-absorption remainder must remain visible whenever the paper leaves the flat-boundary episode. ## Presentation The paper is readable at 45 pages including proofs and references. The abstract and first five pages now convey one mechanism. The result table is useful. All formal statements have proofs, theorem labels are proportionate to their roles, and the PDF has no visible layout failure. Some pages are dense, but moving more formal material out of the main text would save little and could obscure the chain. ## Bottom line The paper has moved from an interesting architecture with several overclaims to a coherent theoretical R&R. Acceptance at a top-five journal would still require the authors to make the conditional equilibrium composition transparent and to close the two interpretation gaps above. Even after those revisions, editorial judgment about the depth and generality of the rotation mechanism will remain the main publication risk.