Round 8: response to reviewers

Protocol version reviewed: 0.8.0. Revised version: 0.9.0.

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. Two corrections to our round 7 responses are disclosed under C8-4 and F8-4.

Summary of main changes

• Cumulative incidence integration uses survival free of both the target and the competing event (formula given); time varying hazards are piecewise constant on digitized reference arm curves.
• Person time is time at risk until the first event of either kind; no rates from median or general follow-up.
• Reference risks: one ordered rule for subdistribution conversions; for cause specific conversions the reference risk comes from the same route as the treated risk.
• O20: primary network exception and conclusions; one input per trial with a stated order; Kaplan-Meier contrast likelihood, standard error and contribution entry.

GPT-6 Astra (gpt-6-astra)

C8-1 (major; Competing risks). Integration used all cause survival instead of survival free of both events.

Response (revised). Agreed; verified (hazards 0.2 and 0.1 per year give 0.518 at five years with event free survival and 0.787 with death only survival). This error entered in version 0.8.0 from wording we adopted. The integration is now written with survival free of both the target and the competing event, using your formula, and grok-4.7 G8-5 flagged the same gloss.

C8-2 (major; Data items; competing risks). Person time fallback permits invalid denominators.

Response (revised). Agreed; your rule adopted: rates use time at risk until the first target event, competing event, censoring or cutoff; the number randomized times mean follow-up only when it is exactly that; median follow-up, treatment exposure and general follow-up are not substitutes; without valid person time no rate is computed by division. grok-4.7 (G8-6) proposed allowing a flagged median; we chose your stricter rule because a median times the number randomized is not time at risk and flagging does not correct the rate.

C8-3 (minor; Absolute effects). Reference risk hierarchy omits integration.

Response (revised). Agreed. Subdistribution conversions take F_ref from the reported Aalen-Johansen incidence, then piecewise integration of reference hazards, then the exponential approximation under supported constant hazards, otherwise relative only; cause specific conversions require hazards, not a reference cumulative incidence alone (see F8-4 for the reference risk shown).

C8-4 (minor; Competing risks; conclusions). O20 exception missing from the primary network rule and conclusions.

Response (revised). Agreed; the primary network sentence now names the O20 exception, and the conclusions rule states that O20 has only its not applicable network, else descriptive reporting. In round 7 our response recorded C7-2 as adopted although the primary network sentence and the one input per trial rule had not been carried (fable-5.1 noted this); both are now in the text.

Fable 5.1 (fable-5.1)

F8-1 (minor; Competing risks). Primary network for O20.

Response (revised). Agreed; see C8-4.

F8-2 (minor; Event outcomes and harms). One input per trial.

Response (revised). Agreed; your order adopted (reported hazard ratio, then cumulative incidence or Kaplan-Meier contrast, then arm counts; alternatives logged).

F8-3 (minor; Competing risks; absolute effects). F_ref and the constant hazard requirement.

Response (revised). Agreed. F_ref follows the stated order, and constant cause specific hazards are required only when F_ref comes from the exponential approximation, which then carries its label.

F8-4 (minor; Absolute effects; competing risks). Two baselines on the cause specific route; pooling of time varying hazards.

Response (revised). Agreed. For cause specific hazard ratios the reference risk shown is the reference node cumulative incidence computed by the same route as the treated risk; time varying hazards are piecewise constant on the intervals of digitized reference arm curves, pooled per interval on the log scale. In round 7 our response to grok-4.7 G7-4 did not disclose that the integral route had not been carried into the absolute effects bullet; it now is.

F8-5 (minor; Contributions; Kaplan-Meier input). Kaplan-Meier contrast missing from the contribution rule; source of its standard error.

Response (revised). Agreed; the contribution rule lists the Kaplan-Meier contrast and the standard error comes from the reported confidence intervals or the Greenwood variance.

Grok 4.7 (grok-4.7)

G8-1 (minor; Known status). Parenthesis sends O20 counts to a subdistribution network.

Response (revised). Agreed; complete status counts are subdistribution inputs only for outcomes with a competing event and not applicable inputs for O20.

G8-2 (minor; Competing risks). Primary network sentence without exception.

Response (revised). Agreed; see C8-4.

G8-3 (minor; Competing risks). F_ref definition against absolute effects.

Response (revised). Agreed; see F8-3 (one definition with the stated order).

G8-4 (minor; Kaplan-Meier contrast; contributions). Likelihood and contribution entry.

Response (revised). Agreed; it enters as a normal likelihood on the difference in log cumulative cause specific hazards and is listed in the contribution rule.

G8-5 (minor; Competing risks). All cause survival gloss is wrong for nonfatal outcomes.

Response (revised). Agreed; see C8-1.

G8-6 (minor; Data items). Person years fallback with a median.

Response (revised (alternative)). Resolved by the stricter rule of gpt-6-astra (C8-2): a median follow-up times the number randomized is not time at risk, so no rate is computed from it; another supported route or the relative only rule applies.