Round 5: response to reviewers

Protocol version reviewed: 0.5.0. Revised version: 0.6.0.

Verdicts: GPT-6 Astra MAJOR REVISION; Fable 5.1 ACCEPT; Grok 4.7 MINOR REVISION. Every comment is answered below and recorded in ../comments.csv. Where reviewers proposed different solutions, the response names the choice and its reason.

Summary of main changes

• Hazard ratio analysis of harms: counts only with known status for every randomized participant; the 95 percent rule as a sensitivity analysis; no claimed bound or direction of bias.
• Competing risks: hazard type recorded for every input and part of the network key; cumulative incidence and complete status counts are subdistribution inputs; conversion formulas; no exemption.
• Recorded harms analysis: no pooled absolute effect; relative range for certainty; per trial display.
• Relative only rule when absolute conversion is unavailable; contribution inputs match the primary model; compatible follow-up restored as 20 to 104 weeks; O22C duplicate and version number fixed.

GPT-6 Astra (gpt-6-astra)

C5-1 (major; Event outcomes and harms). The 95 percent rule lets unknown outcomes become non-events; the bias claim is wrong under unequal loss.

Response (revised). Agreed; verified your example (hazard ratio 1.20 becomes 1.26 with 5 percent lost from one arm). Counts enter the primary hazard ratio analysis only when status is known for every randomized participant at t (with t and known status defined as grok-4.7 proposed); the 95 percent rule is a sensitivity analysis; the protocol now states that it implies no general bound or direction of bias and shows both directions. Counts with unresolved missing status enter the recorded harms analysis.

C5-2 (major; Competing risks). Cumulative incidence contrasts target subdistribution hazards; hazard type missing from the network key; exemption undefined.

Response (revised). Agreed. Estimator and hazard type are recorded for every input including derived ones; cumulative incidence contrasts and complete status counts are subdistribution inputs and never enter cause specific networks; hazard type is in the network key and connectivity rule; the exemption is removed, so the rule applies to every outcome with competing events, including harms; cumulative incidence enters only when 0 < p < 1 with finite uncertainty.

C5-3 (major; Recorded harms analysis). Pooled absolute difference at the median duration not operationally defined.

Response (revised). Agreed. No pooled absolute effect is derived from the recorded harms analysis; each trial recorded proportions and within trial risk difference are displayed with observation period and completeness; magnitude dependent certainty domains use the relative range 1/1.25 to 1.25 on the recorded event rate ratio, with the scale stated. fable-5.1 and grok-4.7 had proposed defining a pooled absolute difference at the median; we followed your position because the observation processes differ between trials, and gave the analysis a fully specified relative rule so that the certainty rules remain executable, which is what their comments required.

C5-4 (minor; Confidence in cumulative evidence). Contribution rule omits cumulative incidence inputs.

Response (revised). Agreed. The contribution network uses, per trial, the input that entered the primary model (reported log hazard ratio, cumulative incidence contrast with its delta method variance, or count derived estimate with the 0.5 rule), preserving shared arm covariance.

C5-5 (minor; Absolute effects; certainty; ranking). Relative only exception incomplete.

Response (revised). Agreed. When valid absolute conversion is unavailable, absolute effects and categories are marked unavailable, magnitude dependent domains are judged against the relative range 1/1.25 to 1.25 with the reason stated (never no concerns merely because conversion is unavailable), and rankings needing absolute effects are omitted; this overrides the general requirement. We used the relative range rather than a purely qualitative judgment so that the rule stays numeric and reproducible.

C5-6 (minor; Composite definitions). Duplicated O22C.

Response (revised). Agreed; corrected.

Fable 5.1 (fable-5.1)

F5-1 (minor; Recorded harms absolute effects). Absolute effects and imprecision undefined for the recorded harms analysis.

Response (revised (alternative)). Resolved differently from your proposal: gpt-6-astra (C5-3) argued that a pooled absolute difference at the median observation period is not defined when observation processes differ, so no pooled absolute effect is derived; per trial proportions are displayed and magnitude dependent domains use the relative range 1/1.25 to 1.25 on the recorded event rate ratio, which makes the certainty rules executable.

F5-2 (minor; Frequentist analysis; certainty; sensitivity list). Compatible follow-up undefined.

Response (revised). Agreed. 20 to 104 weeks restored in both places and added to the sensitivity list.

F5-3 (minor; Competing risks). No primary hazard type.

Response (revised). Agreed. Cause specific is primary for time to event outcomes (outcome trials report Cox models) and subdistribution for binary harms (counts and cumulative incidence estimate it); hazard type is part of the network key.

F5-4 (minor; Competing risks). Negligible mortality undefined.

Response (revised (alternative)). Resolved by removal rather than a cutoff, as gpt-6-astra proposed: because every input now carries its hazard type and counts with complete status are subdistribution inputs, no exemption is needed.

F5-5 (minor; Composite definitions). Duplicated O22C.

Response (revised). Agreed; corrected.

F5-6 (minor; Outcomes and time points). Stale version number.

Response (revised). Agreed; corrected.

F5-note (note; Implementation). Arm level cumulative incidence entry is not a multinma likelihood.

Response (revised). Adopted: cumulative incidence enters as a contrast level normal likelihood on the difference of log(-log(1 - p)) between arms.

Grok 4.7 (grok-4.7)

G5-1 (major; Competing risks; network key). Competing risk synthesis is not yet a network; conversions and exemption.

Response (revised). Agreed. Hazard type in the network key and connectivity rule; your conversion formulas adopted (subdistribution: 1 - (1 - F_ref)^HR; cause specific with both hazards and the stated assumption for the competing hazard); applied to every outcome death can preclude. The exemption is removed rather than set at 1 percent, because counts with complete status are now subdistribution inputs and need none.

G5-2 (major; Hazard ratio analysis). Time and known status undefined for the count rule.

Response (revised). Agreed; your definitions of t and known status adopted. Following gpt-6-astra (C5-1), the primary analysis requires known status for every randomized participant and the 95 percent rule is a sensitivity analysis; your statement that coding unknowns as event free pulls toward the null holds for equal loss, and the protocol now also shows that unequal loss can push away from the null.

G5-3 (minor; Composite definitions). Duplicate O22C.

Response (revised). Agreed; your sentence adopted.

G5-4 (minor; Contributions). Cumulative incidence inputs.

Response (revised). Agreed; see C5-4. Cumulative incidence of 0 or 1 never enters (C5-2), so no correction is needed for it.

G5-5 (minor; Frequentist analysis). Compatible follow-up bounds.

Response (revised). Agreed; see F5-2.

G5-6 (minor; Recorded harms analysis). Absolute differences unnamed.

Response (revised (alternative)). Resolved as in C5-3 and F5-1: the coefficient is named the log relative rate of recorded first events, not a hazard ratio, and certainty uses the relative range on that scale; no pooled absolute difference is derived.