Peer review round 2 of protocol version 0.2.0. Reviewer grok-4.7. The tables are authoritative where they conflict with prose. registry_dose_candidates.csv is an extraction aid, not a node list (Networks and nodes). Its route guesses and titration strings are not treated as nodes.
Round 1 comments
G1. Resolved. Nodes separate original, reformulated, and 25 mg and 50 mg oral semaglutide, and they include the requested doses, routes, and formulations. Emax is limited to one molecule, route, and formulation. Agents with no family row are added by the extraction rule, which is acceptable. False reformulated-semaglutide notes are a new error below.
G2. Resolved. O13A and O13B, O22 and O22B, O24A and O24B, O100A and O100B, and the substance-specific rows match outcomes.csv and are not pooled across those splits.
G3. Resolved. The 80 percent rule and strata S0 to S4 are in populations.csv and the protocol. S4 enters only the pooled patient-important network and a sensitivity analysis.
G4. Resolved. Short trials have windows. Nearest-to-target, the P04 and P05 glycemic window (12 to 40 weeks, target 26, secondary 40 to 64), and the P02 and P03 weight window (40 to 104 weeks, target 68) are an acceptable consistency choice. Overlapping blood pressure windows are a new error.
G5. Partly resolved. The cloglog model with a log follow-up offset, the O04B rate ratio, the descriptive rule below 20 events, and no continuity correction are in the protocol. Replacing 0.1 per 1,000 with a hazard or odds ratio of 1.25 is accepted: at a very low baseline risk that absolute cut is smaller than the events these trials can show. Keeping gastrointestinal events and pancreatitis at participant level is accepted, because episode rates are reported inconsistently. The rare-outcome switch still depends on extracted reference-arm risk (new issues).
G6. Resolved. The 2026-10-12 PubMed test is the right way to settle truncation inside quotes, and R/search_strategy.R matches it: PubMed phrases keep the asterisk inside quotes, while Europe PMC and OpenAlex expand truncation. HRS9531 and HRS-7535 are separate in concepts.yaml, agents.csv, and nodes.csv. Keeping bare cagrilintide is justified by monotherapy arms in CagriSema trials.
G7. Resolved. The reference node is declared before fitting. Horizons are stored, including 156 weeks for O14, O20, O21, O22, O23, and O50, and 52 weeks for O40. Applying 156 weeks inside heart failure populations is a new error.
G8. Partly resolved. Baseline priors N(0, 5^2) on the logit and cloglog scales are accepted: a prior centered at 50 percent with SD 2 puts a risk near 1 per 1,000 far into the tail. Relative priors N(0, 2.5^2), and N(0, 1.5^2) for rarer outcomes, are a reasonable compromise with large weight-loss responder ratios. Turner 2015 priors by outcome type and comparison type are in turner_map.csv and priors.csv. The half-normal prior with scale 1 on τ for nonrandomized evidence is in the protocol. Choosing PR05 versus PR06 from the observed reference risk is a new problem.
G9. Partly resolved. Limiting nonrandomized synthesis to harms is an acceptable scope decision. Mortality, MACE, heart failure, serious adverse events, and discontinuation stay trial outcomes. harms_nrs.csv matches nrs_synthesis = yes. Critical risk of bias is excluded, overlap follows the stated hierarchy, and nonrandomized results are not pooled with trials in the primary analysis. Estimand pooling is a new issue.
G10. Partly resolved. Leaving HFmrEF unsplit is accepted. Dedicated GLP-1 heart failure trials enrolled LVEF at or above 45 or 50 percent, so a 41 to 49 percent network would have no dedicated trials, and LVEF at or above 50 percent remains a sensitivity analysis on P07A. P08A and P08B are separate. The one-versus-two population contradiction is gone.
G11. Partly resolved. CINeMA is applied to both frameworks and the lower rating is kept. Numeric cutpoints and the lock-before-extraction statement are written down. Two rules still use outcome data before that lock. Bayesian contribution weights are unspecified (new issues).
G12. Resolved. Cluster trials use a design effect with the trial ICC or 0.05. A crossover contrast enters only when its standard error is reported and carryover or period is addressed; otherwise the first period is used; otherwise the trial is described. The source hierarchy is ordered, and completers and per-protocol results are out of the primary networks.
G13. Resolved. The checklist, generated Appendix A, CENTRAL, an Embase decision before registration, and quarterly WHO proposed INN checks are in place. Remaining checklist items are administrative.
G14. Resolved. node_classes.csv gives beinaglutide a half-life of minutes.
G15. Resolved. Aspart, lispro, glulisine, and regular insulin are separate nodes, as are repaglinide, nateglinide, and OTHER_ACTIVE. One premixed-insulin node was not required.
G16. Resolved. When every arm has events, the frequentist model is inverse-variance random effects with REML and no continuity correction. Otherwise it is Mantel-Haenszel common effect (Efthimiou 2019). Naming netmetabin is not required for the rule to run.
G17. Resolved. Nonrandomized pooling uses REML, Sidik-Jonkman if REML fails, and Hartung-Knapp intervals.
G18. Partly resolved. The EM02 unit cell names BMI standard deviation score for P03, but the populations column omits P03, so a join does not fit the modifier in the pediatric network.
G19. Partly resolved. Weight and BMI thresholds remain 2.5 percentage points or 2.5 percent, and systolic blood pressure remains 2 mmHg. The reason is accepted as a provisional position: 5 percentage points is a regulatory efficacy benchmark against placebo, and 2 mmHg is tied to population cardiovascular risk. Locking these values is an administrative condition, so the unchanged numbers do not by themselves withhold acceptance.
G20. Resolved. WOMAC is rescaled to 0 to 100 with threshold 10. The eGFR slope is higher-better, with decline coded negative, and the threshold is 0.75 mL/min/1.73 m² per year.
G21. Resolved. O04 is a primary P04 outcome and is windowed. O06 exists for P05 and is windowed with O01. O06 need not be in the P05 primary-outcome list.
G22. Resolved for the pooling ban: O21 is never pooled with O21B. Coverage of P07 and P08 is a new issue.
G23. Resolved. Meta-regression uses one covariate at a time, prior N(0, 1^2) per standard deviation, and a 20 percent missingness rule. Fixed-ratio products are excluded from Emax and enter the component model as present or absent. The extra model of at most two covariates, each in at least 10 trials, does not restore the overfitting design.
G24. Resolved. The OpenAlex run log stores search.title_and_abstract, and Appendix A names that parameter. Calibration requires Cohen's kappa of at least 0.8, which is stricter than 0.70, and reports percent agreement.
New issues
Major
1. Overlapping O05 windows. windows.csv rows W01 and W06 both give O05 in P01 a primary window. W01 is weeks 36 to 64, target 52. W06 is weeks 40 to 104, target 68, populations P02, P01, and P03. Week 52 of systolic blood pressure in P01 has two primary targets. W02 (20 to 36) overlaps W07 (20 to 40), and W03 (64 to 130) overlaps W06 from 64 to 104. Outcomes and time points requires windows for the same outcome and population to be contiguous, with each observation in at most one window. W06 and W07 also list O05 for P03, while outcomes.csv assigns O05 only to P01 and P02.
Replacement: remove O05 from W06 and W07. Keep O05 for P01 on W01, W02, and W03 only. Add a P02-only primary row for O05: target 68, lower 40, upper 104. If a P02 secondary is wanted, add target 26, lower 20, upper 40, population P02 only.
2. False bioequivalence notes. nodes.csv rows SEMAR2_PO_1p5MG_QD, SEMAR2_PO_4MG_QD, and SEMAR2_PO_9MG_QD each say "Bioequivalent to 3; 7 and 14 mg of the original tablet." Read literally, 1.5 mg is pooled with 14 mg. The original-tablet rows correctly pair 1.5 with 3, 4 with 7, and 9 with 14.
Replacement notes: 1.5 mg, "Bioequivalent to original tablet 3 mg. Pooled with SEMA_PO_3MG_QD only in the prespecified sensitivity analysis." 4 mg, the same sentence with 7 mg and SEMA_PO_7MG_QD. 9 mg, the same sentence with 14 mg and SEMA_PO_14MG_QD. In Networks and nodes, replace "pools each with its bioequivalent original dose (3, 7 and 14 mg)" with "pools 1.5 mg with 3 mg, 4 mg with 7 mg, and 9 mg with 14 mg only."
3. One horizon for heart failure and long cardiovascular trials. outcomes.csv sets absolute_horizon_weeks to 156 for O20, O21, and O23, whose populations include P07A and P07B. O40 on those populations is 52 weeks. Heart failure trials last about one year. Under the constant-hazard absolute-effect calculation, a 156 week horizon multiplies a one-year risk difference by about three and can change whether it crosses 5 per 1,000.
Replacement: store population-specific horizons. For O20, O21, and O23 use 156 weeks for P01, P02, P08A, and P08B, and 52 weeks for P07A and P07B. Keep O40 at 52. Split rows or add a population-keyed horizon field.
Minor
4. Priors and the rare-outcome threshold are chosen after extraction. priors.csv PR05 and PR06 switch at a median reference risk of 5 per 1,000. Absolute effects switches the threshold to a hazard or odds ratio of 1.25 when that risk is below 2 per 1,000. Confidence in cumulative evidence says thresholds are locked before any outcome data are extracted. The two cutoffs disagree. The reference risk is not the treatment effect, but the rule cannot be executed as a pre-extraction lock. Several harm rows are already marked relative 1.25, with no rule for a risk above 2 per 1,000.
Replacement: before extraction, assign PR06 and threshold_scale relative to O29, O32A, O32B, O34, O35A, O35B, O36A, O36B, O37A, O37B, O38, and O120. Do not recompute that assignment from the observed median risk. Reference-arm risks may still be used for absolute effects and scenarios, logged before relative effects are modeled.
5. Longest-follow-up windows are not joinable. windows.csv W00 stores prose in outcome_ids, target "longest", and a blank upper bound, so nearest-to-target is undefined. W25 (O51) has the same target and a blank upper bound. O04B (log_rate) is outside W00.
Replacement sentence in Outcomes and time points: "W00 and W25 select the longest randomized follow-up at or above lower_week_inclusive. A blank upper_week_exclusive means no upper limit. The nearest-to-target rule does not apply." Add O04B to that rule. Define W00 membership by model_scale in that sentence, or list the outcome ids.
6. Stratum delimiter. populations.csv P01 contains "S3 prandial; premixed or basal bolus insulin with or without other drugs". The field delimiter is a semicolon, so a split creates a stray stratum. Replacement: "S3 prandial, premixed or basal bolus insulin with or without other drugs".
7. EM02 and children. effect_modifiers.csv EM02 populations omit P03 while the unit cell mentions BMI standard deviation score in P03. Replacement: add P03, and state that the child row uses BMI standard deviation score. Consider adding P04.
8. Expanded composites have no heart failure or CKD row. O21B and O22B are limited to P01 and P02. Replacement: add P07A and P07B to O21B, and P08A and P08B to O22B. Do not pool them with O21, O22, or O50.
9. Nonrandomized estimands. harms_nrs.csv uses "As above" for exposure, comparators, and designs. The O28 estimand cell lists both on-treatment and initial-treatment hazards, and the nonrandomized section pools by measure, design, and comparator type only. Replacement: repeat the exposure, comparator, and design text on each row, and add: "Pool on-treatment and initial-treatment estimates in separate analyses. Do not combine them in one meta-analysis."
10. Relative CINeMA range. Imprecision and incoherence say the discrepancy is "smaller than the threshold" for a ratio of 1.25, with no equivalence range. Replacement: for relative thresholds the range is 1/1.25 to 1.25, and the discrepancy is the ratio of the two ratios, or the difference of log effects compared with log 1.25.
11. Blank horizon. O120 has no absolute_horizon_weeks, while Absolute effects assigns every binary outcome an absolute effect at that horizon. Replacement: set 52, or exempt nonrandomized-only outcomes in that sentence.
12. Reference-node frequency. "The most frequent active node" does not define the count. Replacement: "the active node that appears in the most trials."
13. Flexible oral semaglutide. SEMA_PO_FLEX_QD has a blank formulation. The note cites PIONEER 7, which used the original tablet. Replacement: set formulation to "original tablet". Add a reformulated flexible node only if such a trial exists, by the extraction rule.
14. Dose templates. Family rows such as ECNO_SC use a literal dose token. If that row is used as a node, doses collapse. Replacement: mark these rows as templates excluded from every network until a numeric dose replaces them.
15. Bayesian contribution weights. Frequentist CINeMA uses netcontrib, shortest path. No Bayesian weight is defined. Replacement: use that same shortest-path graph for the Bayesian estimate, or apply contribution domains only to the frequentist estimate and use the Bayesian estimate for the effect and for imprecision. Keep the lower rating when both are rated.
Administrative conditions
Team names, affiliations, funding, competing interests, Embase access, PRESS review by a human information specialist, verification of Chinese terms by a Chinese-reading team member, and confirmation and locking of the provisional thresholds are pre-registration conditions. Their absence does not by itself withhold acceptance. The methods are fit to register once those conditions are completed and the three major items are corrected. Those items are local table and sentence fixes. They are wrong or non-executable if the protocol is applied as written, so acceptance waits on the correction.
VERDICT: MINOR REVISION