# Independent Referee Report on v3.7 **Manuscript:** *The Task Frontier: Horizon, Forgiveness, and the Direction of AI Innovation* **Version reviewed:** v3.7, 38 pages, 2026-07-07 **Review date:** 2026-07-14 **Review basis:** full source and proof appendix; existing reports were consulted only after the independent read; the 95-check companion script was rerun; current primary sources were checked through 2026-07-14. ## 1. Recommendation ### Top-five general-interest journal **Reject in present form, with encouragement to rebuild and resubmit elsewhere.** The paper has a top-five-caliber question and a potentially distinctive mechanism, but the current formal architecture does not yet establish several of the claims advertised in the abstract. My assessment of an R&R in present form is in the single digits to low teens. The best eventual fits are AER or Review of Economic Studies if the equilibrium and measurement problems below are solved. The paper is currently too conditional for Econometrica and too lightly disciplined for the broad claims expected at QJE or JPE. ### Strong field journal **Major revision / reject-and-resubmit.** The two-dimensional task idea and the certification mechanism could support a strong RAND or AEJ: Micro paper, but the exact versus local inconsistencies should be repaired before submission even there. This is not a verdict against the protected messages. It is a verdict that the present theorems do not yet prove them at the advertised level of generality. ## 2. What is genuinely promising The paper asks an important economic question: why does rapid commoditization of known AI capabilities coexist with escalating frontier investment, and why should investment move in one technical direction rather than another? Its best insight is that appropriability may be a property of a task's statistical structure. A demonstrated plan can be observed, while a low tail-failure rate takes evidence to establish. Mapping that asymmetry into directed innovation is potentially original and economically important. The manuscript is unusually readable for a theory paper. It has a clear puzzle, strong examples, a coherent vocabulary, and a useful attempt to connect the model to observable lags. The proof and verification infrastructure is also valuable. These assets should be preserved. ## 3. Most damaging objections ### 1. The claimed exact two-dimensional reduction is not the probabilistic model used downstream This is a foundational internal inconsistency. Lemma 1 correctly derives that retries change two objects: 1. the reliability threshold; and 2. the *scale* of the Gumbel success kernel, which is `1/(1+ell)`. The manuscript then reduces a primitive task `(n, phi, ell, epsilon)` to `(h,f)` and writes aggregate surplus using a common kernel `P_F(q_F-f)`. That reduction discards `ell`, even though `ell` changes the slope of the success curve. Two primitive tasks can have the same `(h,f)` and different `ell`, yet have different success probabilities at every capability away from the threshold. A direct counterexample using the paper's exact technology, with `n=100`, `epsilon=.1`, and the same `(h,f)=(ln 100,0)`, gives success at `q_F=1` of 0.962 for `ell=0`, 0.9983 for `ell=4`, and 0.9998 for `ell=9`. Relatedly, the Gumbel *location* in Lemma 1(ii) is `phi + ln(n)/(1+ell)`. The fragility threshold for tolerated failure `epsilon` is approximately `phi + {ln(n)+ln(1/epsilon)}/(1+ell)`. Calling both objects the same location obscures a quantile shift. **Why it matters:** `W`, `B_H`, `B_F`, `C`, the profit-gap representation, rotation, and welfare all use the aggregate probabilistic model. The capability frontier can still be two-dimensional, but the task measure must retain the task-specific kernel or the main model must explicitly use deterministic service thresholds. **Repair:** integrate over primitive task types (or over `(h,f,ell)`) while keeping the system capability vector two-dimensional. Define task-specific reliability kernels. The boundary and supermodularity results survive; the claimed lossless reduction of a probabilistic task to `(h,f)` does not. ### 2. The scalar-representation corollaries are false as written The core order-theoretic theorem—scalar representation iff requirement pairs form a chain—is correct for the deterministic service relation. Its advertised equivalence to homogeneous forgiveness is not. Theorem 1(i) states that common `(ell,phi)` is sufficient for a chain, but the proof quietly adds common `epsilon`. Without common tolerated failure, the statement is false. For example, with common `ell=phi=0`: - a 100-step task with `epsilon=.5` has `(h,f)=(4.605,4.968)`; - a 1-step task with `epsilon=10^-12` has `(h,f)=(0,27.631)`. The pair is unordered. Conversely, heterogeneous forgiveness can still generate a chain. Therefore “a scalar index is exactly the case in which all tasks forgive alike” is stronger than the theorem and not true without restrictions on local difficulty and tolerated failure. **Repair:** keep “chain iff scalar” as the theorem. State common `(ell,phi,epsilon)` as one sufficient benchmark. State a support condition under which heterogeneous forgiveness produces unordered pairs. This is a precision repair, not a change in the paper's economic message. ### 3. The exact welfare wedge and its numerical calibration do not follow from the annuity theorem Theorem 3 says private marginal value is a finite lag-window integral *plus* a remainder `R_i`, and its annuity form is an approximation with an explicit boundary-variation error. With rising reliability lags, the proof adds another positive infinite-horizon term. Theorem 6 then drops all of these terms and states the exact relative wedge as `(1-exp(-r tau_H))/(1-exp(-r tau_F))`. That equality holds only in the locally flat/stationary benchmark. In a model whose central result is that boundary values change enough to cause rotation, flat boundary values cannot silently be treated as the global baseline. The calibrated claim that the market values horizon at exactly 0.30 of reliability is therefore not established by the general model. **Repair:** derive the exact private and social marginal-value integrals first. Then either (a) prove bounds on the direction wedge using the existing variation terms, or (b) state a separate local-stationary proposition with its assumptions in the formal statement and label 0.30 a local illustration. The exact theorem and the calibration must refer to the same object. ### 4. Rotation is still a conditional path result, not a dynamic equilibrium result The manuscript now acknowledges this, which is an improvement, but the abstract and conclusion continue to say that investment rotates. The proof assumes a pure horizon push, holds `q_F` and the rotation threshold fixed, and shows that if the push continues through depletion, the scarcity ratio crosses once. It does not show that an optimal policy follows that path. The calendar-time comparative static is also not proved generally. A shorter horizon lag lowers the crossing threshold, but it also lowers the return to horizon investment and can slow the approach to that threshold. A lower crossing *position* need not mean an earlier crossing *time* under convex costs. **Repair:** solve a tractable optimal-control problem and establish a switching region, or restrict the formal claim to a state-space regime boundary. Because rotation is a protected headline message, the top-five route is the former. A parametric task density is acceptable if it produces an actual optimal switching theorem and transparent comparative statics. ### 5. The certification floor is real, but the imitation lag is not yet an equilibrium object The Bernoulli testing lower bound is valid. Several economic conclusions are stronger than that bound: - Part (iii), “achieving is as hard as certifying,” is a minimax two-point statement, not a universal statement about invention. The identity of the reliable policy must be unknown and the available information restricted to the modeled trials. - Part (iv) says reliability appropriability is strictly larger for any discount rate, but an increasing lower bound on `tau_F` does not imply `tau_F>tau_H` at every state. That inequality needs to be assumed or derived beyond a frontier threshold. - The data-moat proposition divides a common sample target by assumed flows. It does not model who supplies verification, who pays for it, strategic disclosure, endogenous testing effort, correlation, or cross-task generalization. - `D_pub` is allowed to be zero even though the displayed lag ratio divides by it. - The simulator result is task-specific. A general simulator or formal guarantee may amortize evidence across tasks; the present “every route” language is too broad. **Repair:** add a certification-and-entry game. Buyers or a regulator require a certificate; firms choose evaluation effort or wait for public evidence; deployment creates proprietary evidence; and cross-task generalization is parameterized. Derive the entrant's equilibrium delay and the conditions under which it exceeds the plan lag. The statistical lower bound should discipline that game, not substitute for it. ### 6. The herding result is a local 2x2 game but is repeatedly interpreted as an industry dynamic The local payoff comparison is clear, but it is not a repeated equilibrium. It assumes that the rival's next same-size breakthrough matches a lead and that direction is chosen once. The “separability to first order” remark is an honest local argument; it does not justify global portfolio convergence. There are also two formal problems: - At equality in the herding condition, reliability is only weakly dominant and the herding equilibrium is not unique; the theorem needs a strict inequality for uniqueness. - The cheap-intensity limit lets sole-occupant value grow without bound while the paper's `W` is bounded. It extrapolates a constant local boundary value through infinitely many innovations, outside the domain of the approximation. Thus the claimed continuous-progress limit is not a result of the full model. **Repair:** either solve a small repeated/Markov direction game, or state herding as a local strategic bias and remove the global and infinite-intensity claims. Since herding is protected, a finite-state Markov version is the stronger route. ### 7. The policy headline does not derive market concentration or unambiguously safer AI The corollary defines the capability gap as the “moat” and then equates a larger gap with greater concentration. The product-market model does not derive that mapping. With equal costs and componentwise dominance, the leader already wins every task for any positive gap; a larger gap raises margins, but not necessarily market share or HHI. Likewise, public verification shortens the fringe lag but weakens frontier investment. It diffuses a given capability more quickly; it does not necessarily raise the absolute reliability of the best or fringe system at every future date once the investment response is included. “Both policies deliver safer AI” is therefore not established. **Repair:** define a market-structure statistic and derive it from heterogeneous costs, task-specific runner-ups, or entry. Separate three outcomes: frontier reliability, diffused reliability, and concentration. The opposite-sign competitive-advantage result can survive, but safety and concentration require explicit conditions. ### 8. The three aggregate facts are not current enough to carry their assigned weight The public-data audit changes the evidentiary picture: - METR's current Time Horizon 1.1 data still support rapid growth, with a stitched all-time 50% doubling estimate around 188 days, but METR documented a March 2026 regularization error and substantial sensitivity of 80% horizons to modeling choices. Current frontier 50%/80% ratios vary materially across models rather than sitting at a stable value near five. The ratio should not be a “fact” without a dated dataset and uncertainty calculation. - Epoch's May 2026 analysis supports an aggregate open-weight lag of about four months. It does not provide the paper's high-reliability lag of 14 months. - METR's DeepSeek evidence supports roughly a six-month autonomy comparison in a particular evaluation. It is not an estimate of the model's task-specific certification lag. - The inference-price decline is well sourced. Rising frontier spending needs its own current source rather than appearing as the uncited half of Fact 3. **Repair:** rebuild the aggregate section from dated, reproducible data. If a defensible `tau_F` cannot be estimated, remove the 14-month point calibration and present the lag ratio as the paper's key missing measurement target. ## 4. Result-by-result assessment | Result | Assessment | Reason | |---|---|---| | Lemma 1 | Useful derivation, but misused downstream | Threshold and limit are sound; heterogeneous kernel scale is discarded later. | | Theorem 1 | Correct core, false corollaries | Chain iff scalar is elementary and useful; homogeneous-forgiveness equivalence is overstated. | | Lemma 2 | Standard benchmark | Clean gap identity under restrictive runner-up/equal-cost assumptions; outside option and equilibrium selection should be explicit. | | Theorem 2 | Promising mechanism, overgeneralized economics | Testing lower bound is standard; equilibrium lag and universal invention claim do not follow automatically. | | Proposition 1 | Arithmetic benchmark | Flow-ratio result is assumed into trial flows; simulator statement needs scope and domains. | | Theorem 3 | Useful accounting decomposition | Exact formula includes a potentially large infinite-horizon remainder; “annuity pricing” is local. | | Lemma 3 | Correct with a strictness repair | `>=` the mode gives only weak decline unless positive mass lies strictly past it. | | Theorem 4 | Conditional comparative static | Shows crossing along a maintained pure push, not an optimal switching path. | | Theorem 5 | Clean local game | Standard 2x2 logic; not a dynamic equilibrium; equality and global-limit claims need repair. | | Proposition 2 | Local race calculation | Own/rival intensities are tractable, but the unbounded sole value contradicts bounded total surplus outside the local approximation. | | Theorem 6 | False as a general exact theorem | Drops Theorem 3's remainder, variation, and rising-lag term. Valid as a local-flat benchmark. | | Corollary 1 | Suggestive, not market structure | Gap comparative statics are algebraically correct; concentration and safety implications are not derived. | ## 5. Novelty and literature audit The precise combination of retries, tail certification, imitation lags, and directed innovation still appears distinctive. The novelty claim is not yet secure because the paper omits or under-engages several close literatures: - Hagiu and Wright, “Data-enabled learning, network effects, and competitive advantage,” RAND 2023, directly models dynamic competition when customer data improve products and studies data-sharing policy: https://doi.org/10.1111/1756-2171.12453 - DeMarzo, Kremer, and Skrzypacz, “Test Design and Minimum Standards,” AER 2019, provides a formal certification environment rather than treating certification as a fixed gate: https://doi.org/10.1257/aer.20171722 - Gans, “Artificial Jagged Intelligence,” NBER 2026, studies task-distribution mismatch, reliability investment, and verification: https://www.nber.org/papers/w34712 - Gans and Goldfarb, “O-Ring Automation,” NBER 2026, is appropriately cited but needs a sharper distinction on complementarity and task aggregation: https://www.nber.org/papers/w34639 - Xu, Wang, Chen, and Xie, “The Economics of AI Foundation Models,” 2025, contains an AI data-flywheel mechanism and openness policy: https://arxiv.org/abs/2510.15200 - Tong and Manea, “Data Sharing and Competition in Learning-by-Deploying Industries,” June 2026, directly studies deployment-generated learning and data pooling: https://arxiv.org/abs/2607.00168 - Korinek and Stiglitz, “Steering Technological Progress,” NBER 2026, is now the current broad reference for policy that directs technical change: https://www.nber.org/papers/w34994 - Bergemann, Bonatti, and Smolin, “The Economics of Large Language Models,” 2025, includes high-dimensional task requirements and error sensitivity: https://arxiv.org/abs/2502.07736 The paper can distinguish itself: none of these sources, based on the current audit, combines a task-specific statistical certification floor with a direction-specific appropriation wedge. But that distinction becomes a top-five contribution only after the certification lag and direction choice are closed in equilibrium. ## 6. Current-data audit Primary sources checked: - METR Time Horizon 1.1 and raw data: https://metr.org/time-horizons/ - METR methodological sensitivity and the 2026 regularization correction: https://metr.org/notes/2026-03-20-impact-of-modelling-assumptions-on-time-horizon-results/ - Epoch's May 2026 open/closed capability lag: https://epoch.ai/data-insights/open-closed-eci-gap - Epoch's inference-price trends: https://epoch.ai/data-insights/llm-inference-price-trends - Epoch's current company and compute-spending data: https://epoch.ai/data/ai-companies ## 7. Top-five scorecard | Dimension | Current score (5=max) | Assessment | |---|---:|---| | Importance of question | 4.5 | Large and timely economic puzzle. | | Distinctiveness of mechanism | 3.5 | Promising synthesis; equilibrium closure needed. | | Theorem depth | 2.0 | Too many identities/local comparisons carry theorem labels. | | Internal model coherence | 1.5 | Task reduction and downstream kernel conflict. | | Proof correctness / scope | 2.0 | Several statements exceed their proofs. | | Empirical discipline | 1.5 | Key reliability lag is not measured; one fact is method-sensitive. | | Exposition | 4.0 | Clear, memorable, and well organized. | | Current top-five readiness | 2.0 | High-upside major rebuild, not submission-ready. | ## 8. Minimal path to a credible top-five submission 1. Repair the task environment and representation theorem before touching prose. 2. Turn certification delay into an equilibrium outcome with endogenous evidence acquisition and explicit deployment gating. 3. Prove an actual optimal switching result for rotation in a tractable dynamic model. 4. Replace the one-shot herding game with at least a small repeated/Markov game, or sharply localize every claim. 5. Restate the welfare wedge as an exact integral result plus a local annuity corollary; derive market structure rather than naming a gap “concentration.” 6. Rebuild the aggregate evidence using the current METR/Epoch data and eliminate the unsupported 14-month calibration unless a reproducible estimate can be produced. 7. Reassess the revised paper from a clean PDF, with the current report hidden from the reviewing pass. ## 9. Bottom line The paper should not be submitted to a top-five journal in v3.7 form. It should also not be abandoned or cosmetically shortened. The core idea is better than the current formal execution. A disciplined rebuild can preserve every protected message while replacing the conditional and approximate links with the equilibrium results those messages need.