Peer review round 15: LivingGLP1 protocol 0.15.0 (reviewer grok-4.7)
1. Round 14 comments
G14-1. Resolved. With one covering arm, (\tau_k) is not estimated and (u_{\mathrm{new}}) is not drawn. Each draw of the reference log hazards is a draw from (\mathrm{N}(y_a, V_a)). With two or more covering arms, the draw is (\mu + Z u_{\mathrm{new}}) at that draw's (\tau_k). The next sentence exponentiates those log hazards and integrates them with the formulas for (S_a) and (F_{1a}). The displayed reference cumulative incidence remains the median of the integrated draws. Scenario draws from (\mathrm{N}(y_a, V_a)) are exponentiated before integration. The one arm predictive draw and the hazard scale now match.
G14-2. Resolved. After the zero count sentence, (y_a) is the vector of estimated log hazards, one coordinate per event and common interval, in that order. For counts, (V_a) is diagonal with the variances given in the previous sentence, approximately (1/d_k), so the zero covariance already stated is the off diagonal of that matrix. Curve inversion still sets (V_a = J\Sigma J^{\mathsf{T}}) for the same vector of log hazards. The pooling likelihood can use (y_a) and (V_a) for both inputs.
G14-3. Partly resolved. The case this comment named is fixed. When the identical reported boundaries are the single interval from 0 to the analyzed horizon, that piecewise fit is the exponential approximation, it is used only when constant hazards are supported, and it is labeled as such. (d_k) is the count of event (k) from time 0 to the analyzed horizon, (\lambda_k) is the exponential of the corresponding log hazard, and a zero count makes the route unavailable.
The same bullet then adds a clause that was not in the adopted text: the separate route's totals meet the identical boundary requirement by construction, and the same pooling model applies. That clause is unsound. Piecewise integration is available only when reported boundaries are identical. A route that may be used only when piecewise integration is unavailable cannot, in the same sentence, satisfy that requirement. The support judgment and the label sit only in the first sentence, so this second route can be read as a constant hazard model without either. The overlap G14-3 and F14-1 closed for a shared reported interval is open again for the constructed interval.
Replace the exponential approximation bullet with: "If every covering arm's reported intervals through the analyzed horizon are the single interval from 0 to that horizon, that piecewise fit is the exponential approximation: it is used only when constant hazards are supported, and it is labeled as such. The separate exponential route is used only when those reported boundaries are not identical, so the piecewise route is unavailable, and every covering arm still supplies both event counts and person time from 0 to the analyzed horizon. Those totals define one constructed interval from 0 to the horizon. The constructed interval does not make the piecewise route available. The same pooling model applies. This route is used only when constant hazards are supported, and it is labeled as such. Here (d_k) is the count of event (k) from time 0 to the analyzed horizon (variance approximately (1/d_k)), the rate (\lambda_k) is the exponential of the corresponding log hazard, a zero count for either event makes the route unavailable, and (F(t) = \lambda_1/(\lambda_1+\lambda_2),(1-e^{-(\lambda_1+\lambda_2)t})) is computed for each draw, with (t) in years."
2. New issues
No major errors.
Minor 1. The separate exponential route contradicts its own availability rule. File: protocol/protocol.qmd, Competing risks, Exponential approximation. This is the residue of G14-3 above. The replacement in that item is the fix.
Minor 2. The new common scale for Kaplan-Meier values and rates leaves (t) and the variance input unspecified. File: protocol/protocol.qmd, Absolute effects, Hazard based outcomes.
F14-4 is the right model. For a constant hazard, (\log{-\log(1-F)}) equals (\log(\lambda t)). I checked that identity at (\lambda = 0.05) per year and (t = 1) year: both equal (\log(0.05)). PR17 is the right heterogeneity prior, and the parenthetical added for F14-2 correctly reserves PR16 for the competing risk routes.
Two readings of the new sentence are executable and they do not agree. The analyzed horizon in this bullet is stated in weeks (about 52 weeks for heart failure networks). Rates elsewhere are events per person year, and the hazard formulas use time in years. If (t) is 52, a rate of 0.05 per year becomes a cumulative incidence of about 0.93. If (t) is 1 year, the same rate becomes about 0.049. The variance phrase says "from its reported interval." In Competing risks, a reported interval is a time partition, which is not a delta method input. The Kaplan-Meier contrast earlier in Competing risks already uses a delta method standard error from the reported confidence interval or the Greenwood variance.
Replace "both enter one model on the log cumulative hazard at the horizon, (\log{-\log(1 - F)}) for a Kaplan-Meier value with variance by the delta method from its reported interval, and (\log(\lambda t)) for a rate with variance approximately (1/d), under PR17" with: "both enter one model on the log cumulative hazard at the horizon under PR17. A Kaplan-Meier value contributes (\log{-\log(1-F)}), with variance by the delta method from its reported confidence interval or from the Greenwood variance, requiring (0 < F < 1). A rate contributes (\log(\lambda t)), where (\lambda) is the event rate per person year and (t) is the analyzed horizon in years, with variance approximately (1/d)."
3. Checks and registration
Against the version 0.14.0 snapshot, the only authoritative table that changed is amendments.csv (A014: 2026-10-13, version 0.15.0, Competing risks and absolute effects, eight comments, before results). comments.csv has those eight round 14 rows. priors.csv is unchanged: PR15 on the reference log hazard, PR16 on (\tau_k) for the competing risk routes, PR17 on the other reference arm risks. PR16 at Half-normal((0, 1^2)), with Half-normal((0, 0.5^2)) and Half-normal((0, 2^2)) as sensitivity priors, matches the reason given for leaving F14-note on the prior scale. Keeping reported Aalen-Johansen values that miss the horizon for the pre-analysis amendment is the deferral already agreed in round 12. outcomes.csv has 77 outcomes. The subtitle, the status sentence, and the outcomes section all say version 0.15.0, dated 2026-10-13, after fourteen rounds. No 0.14.0 string remains in protocol.qmd.
search/strategies/concepts.yaml remains version 1.2.0. R/search_strategy.R builds the PubMed, Europe PMC (excluding patents), OpenAlex, and ClinicalTrials.gov queries from the class terms, agent terms, mechanism block, and PubMed vocabulary. Chinese terms stay out of those functions and remain the manual CNKI and Wanfang list. The renderer was read and was not executed. The iGlarLixi switching row remains a template (is_template yes) and stays out of every network until a strategy is recorded. The Aalen-Johansen inversion formulas are unchanged. Under constant hazards 0.04 and 0.02 per year for one year they still return those hazards. The Quarto file was read and was not rendered.
Team names, affiliations, 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 administrative pre-registration conditions. They do not determine this verdict. The one arm draw, the count assignments for (y_a) and (V_a), the gate on a single reported interval, the PR16 cross reference, and the shared log cumulative hazard scale are in place. The separate route's precondition, and the unit and variance wording of the new scale, still need the replacements above before registration.
VERDICT: MINOR REVISION