Round 4: response to reviewers

Protocol version reviewed: 0.4.0. Revised version: 0.5.0.

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.

Summary of main changes

• Harms: a hazard ratio analysis (reported hazard ratios, cumulative incidence, and counts only with at least 95 percent known event status) and a separately rated recorded harms analysis; attenuation from incomplete observation quantified.
• Competing risks: cause specific and subdistribution hazard ratios separated; cumulative incidence for absolute risks.
• CINeMA contributions: same evidence as the primary model, 0.5 for zero cells in contributions only, qualitative fallback.
• Kidney composite precedence; definition stratum in the network key; iGlarLixi switching template; denominators; complete sensitivity list.

GPT-6 Astra (gpt-6-astra)

C4-1 (major; Event outcomes and harms). Counts with incomplete observation must not be interpreted as hazard ratios.

Response (revised). Agreed; your structure is adopted. In round 4 fable-5.1 and grok-4.7 judged the previous decline sound, and the change keeps what they required (no odds ratios pooled across very different durations): a hazard ratio analysis uses reported hazard ratios, reported cumulative incidence at a common time, and counts only when event status is known for at least 95 percent of randomized participants in every arm (100 percent in a sensitivity analysis), with the attenuation from incomplete observation quantified in the protocol (true hazard ratio 2.0 with risks 0.50 and 0.75 gives 1.93, 1.88 and 1.79 with 5, 10 and 20 percent lost; under 2 percent bias for risks below 5 percent); a separately rated recorded harms analysis of all trials has no hazard ratio interpretation and no fixed horizon absolute risks. Conclusions use the hazard ratio analysis when it meets the connectivity rule.

C4-2 (major; Confidence in cumulative evidence). Count derived contribution networks fail with zero cells.

Response (revised). Agreed. The contribution network uses the same trial level log hazard ratios as the primary model and adds 0.5 to every cell of a trial with a zero event or all event arm, for contributions only (they depend on precision weights and network geometry, and the correction never enters an effect estimate); every primary trial is kept; when valid contributions still cannot be computed, within study bias and indirectness are judged qualitatively across the full evidence set with reasons recorded. This is the zero cell rule with justification you asked for, and the rule fable-5.1 and grok-4.7 proposed.

C4-3 (major; Absolute effects). Competing risks unaddressed.

Response (revised). Agreed. For outcomes that death can preclude, cause specific and subdistribution hazard ratios are extracted and synthesized separately; absolute risks use cumulative incidence accounting for competing deaths, never one minus Kaplan-Meier; cause specific hazard ratios are converted only with the competing event hazard, otherwise relative effects only; trials with negligible mortality are exempt.

C4-4 (minor; nodes.csv iGlarLixi switching). Switching row lacks template protection.

Response (revised). Agreed. IGLARLIXI_SWITCH_<strategy>_SC_TIT_QD with is_template = yes and the proposed note; the text adds this third template category.

Fable 5.1 (fable-5.1)

F4-1 (minor; Event outcomes and harms). No likelihood route for Kaplan-Meier input.

Response (revised). Agreed. Reported cumulative incidence enters on the log cumulative hazard scale with a delta method standard error and offset log t, as proposed; for outcomes with competing deaths the cumulative incidence must account for them (C4-3).

F4-2 (minor; Composite definitions; O04). Hypoglycemia mapping versus the composite rule.

Response (revised). Agreed. Named as a prespecified exception, flagged and excluded in a sensitivity analysis.

F4-3 (minor; Confidence in cumulative evidence). Zero event arms break the contribution network.

Response (revised). Agreed; see C4-2.

F4-4 (minor; outcomes.csv O22E). Populations and overlap with O22D.

Response (revised). Agreed. P08A and P08B added; composites with cardiovascular death are O22D (precedence rule in O22D, O22E and O50).

F4-5 (minor; Sensitivity analyses). Missing two analyses.

Response (revised). Agreed. The list now includes iGlarLixi node combination, the follow-up completeness exclusion, the 100 percent known status restriction and the O04 mapping exclusion.

Grok 4.7 (grok-4.7)

G4-1 (major; Confidence in cumulative evidence). Hazard scale contribution matrices do not run on the primary evidence.

Response (revised). Agreed; see C4-2 (your wording for the network).

G4-2 (minor; outcomes.csv O22D O22E O50). Kidney composite precedence.

Response (revised). Agreed; your three mapping rules adopted verbatim and the Composite definitions text updated.

G4-3 (minor; Networks and nodes; connectivity). Network key omits definition stratum.

Response (revised). Agreed; definition stratum added to the network key and the connectivity rule.

G4-4 (minor; Event outcomes and harms). Kaplan-Meier input and denominators.

Response (revised). Agreed. See F4-1; denominators are the number randomized, with the safety population only when the randomized number is not reported, flagged.

G4-5 (minor; nodes.csv iGlarLixi switching). Switching node.

Response (revised). Agreed; see C4-4.

G4-6 (minor; Sensitivity analyses). Missing analyses.

Response (revised). Agreed; see F4-5.

G4-7 (minor; Absolute effects). One covering reference arm.

Response (revised). Agreed; that arm risk is used and no between arm heterogeneity is estimated.