Peer review round 12: LivingGLP1 protocol 0.12.0 (reviewer grok-4.7)
1. Round 11 comments
G11-1. Partly resolved. The Kaplan-Meier condition (0 < S_k(t_e) < S_k(t_s)) for each event, count intervals that use the reported counts without a number at risk, the exponential definition of (d_k), the data items wording for the two curve pairs, and one shared draw for a subdivided interval are all in the text.
Declining the reconstructed count equations is right. Censoring inside an interval can move without changing the Aalen-Johansen curves, so the event count is not a function of the increment and the number at risk. In the round 11 histories, 100 participants end at the same (F_1 = F_2 = 0.10) with either 5 or 10 events of each type. A known count is still the wrong variance for a curve inverted log hazard: with 100 people and 90 events before any censoring, Greenwood and the delta method give (\mathrm{Var}{\log(-\log S)} \approx 0.01698), while (1/90 = 0.01111). A reported or author supplied joint covariance, mapped by the Jacobian, is the right replacement. The Jacobian now written is the derivative of (h_k), so the variance half of G11-1 is still open (Major 1).
G11-2. Resolved. For a competing event outcome, each covering arm's scenario risk is its own cumulative incidence at the analyzed horizon by the route in use: the reported Aalen-Johansen value when that value is the source, otherwise the cumulative incidence from that arm's own qualifying hazards. A reported incidence is a complete scenario source.
2. New issues
Major 1. The curve covariance equals (J\Sigma J^{\mathsf{T}}) for the wrong function, and per interval pooling drops the shared draw. File: protocol/protocol.qmd, Competing risks, Piecewise integration, from "Uncertainty: for reported first" through "median of the predictive distribution."
The inversion formulas return hazards. For a Kaplan-Meier curve, (h_k = -\log{S_k(t_e)/S_k(t_s)}/\Delta). The log hazard is (\log h_k). The text calls (J\Sigma J^{\mathsf{T}}) the covariance of the log hazards and defines (J) as the Jacobian of those formulas, which is (\partial h/\partial\theta). That product is (\mathrm{Var}(h)). (\mathrm{Var}(\log h) = \mathrm{Var}(h)/h^2). The same formulas recover constant hazards 0.04 and 0.02 per year from the matching cumulative incidences, and at those values the ratio is 625 and 2,500. Log hazards are what get pooled. A literal Jacobian mis-weights the reference arms by those factors and shifts the absolute risk intervals used for the thresholds and the certainty ratings.
The same paragraph requires this dependence to be preserved, and says a reported interval is carried onto the grid as one draw. It then pools log hazards separately on each grid interval with the univariate random effects model. Independent predictive draws cut the delta method variance of the cumulative hazard. One log hazard draw of variance (v), copied onto (n) equal grid pieces, has cumulative hazard variance proportional to (n^2 v). Independent draws of those pieces have variance proportional to (n v). A two way split understates that variance by half. At (n = 10) the factor is 10.
Replace that passage with: "Uncertainty: for reported first event counts with valid person time, the log hazards of the two events are independent, with (\mathrm{Var}(\log h_k) = 1/d_k) and (d_k) the reported count of event (k). A zero count for either event makes the route unavailable. For curve inversion, a reported or author supplied covariance (\Sigma) of the published ordinates that enter the formulas is required, for both events and every interval end that is used, including the analyzed horizon when that ordinate is used. When event free survival is published instead of the competing cumulative incidence, the ordinates are the target cumulative incidence and the event free survival. Let (g) map those ordinates to the vector of log hazards (\log h_k) on the reported intervals. The within arm covariance of that vector is (J\Sigma J^{\mathsf{T}}), where (J) is the Jacobian of (g). Copies of one reported interval are one coordinate of the vector. These vectors are pooled across covering arms in one multivariate random effects model of the reference risks. The reference hazard on each grid interval is the median of its marginal predictive distribution. Uncertainty of the integrated cumulative incidence uses draws from the joint predictive distribution, so copied intervals stay one draw. When any covering arm lacks qualifying inputs, positive finite hazards, or a finite covariance of this form, this route is unavailable and another eligible route is kept. No constants are added and no interval or arm is omitted."
No minor errors were found.
3. Checks and registration
The authoritative tables match the version 0.11.0 snapshot except protocol/amendments.csv row A011 (2026-10-13, version 0.12.0, before results). outcomes.csv has 77 outcomes. The outcomes section and the subtitle both say version 0.12.0, dated 2026-10-13, after eleven rounds. search/strategies/concepts.yaml remains version 1.2.0. R/search_strategy.R still builds the PubMed, Europe PMC, OpenAlex, and ClinicalTrials.gov queries from the class terms, agent terms, mechanism block, and PubMed vocabulary. chinese_terms stays out of those functions and remains the manual CNKI and Wanfang list. The iGlarLixi switching row remains a template (is_template yes) and is excluded from every network until a strategy is recorded. The renderer was read and was not executed. The hazard point estimates are unchanged and still recover 0.04 and 0.02 per year under the stated constancy.
Team names, affiliations, postal address, funding, competing interests, confirmation of Embase access, PRESS review by a human information specialist, verification of the Chinese terms by a Chinese-reading team member, and confirmation and locking of provisional thresholds by clinical advisors and patient partners remain conditions on the administrative checklist. This verdict is based on the methods. The covariance route is the right replacement for reconstructed counts. The Jacobian has to be the derivative of (\log h_k), and the grid copies have to stay one draw in the integrated variance, before registration. With that passage corrected, the methods are acceptable for registration once the administrative conditions are completed.
VERDICT: MAJOR REVISION