Round 10: response to reviewers

Protocol version reviewed: 0.10.0. Revised version: 0.11.0.

Verdicts: GPT-6 Astra MAJOR REVISION; Fable 5.1 MINOR REVISION; Grok 4.7 MINOR REVISION. All three identified the same incorrect hazard derivation, which we introduced in version 0.10.0. Every comment is answered below and recorded in ../comments.csv.

Summary of main changes

• Exact inversion of curves to piecewise constant cause specific hazards (Aalen-Johansen and Kaplan-Meier forms), reporting intervals defined by published numbers at risk, variance 1/d_k, and unavailability rules.
• Only first event curves qualify; mortality curves that include deaths after a nonfatal target event do not.
• Exponential approximation pooled across arms with the same variance and zero count rules; failure of its support source no longer blocks other valid routes.
• Risk scenarios under competing risks use each arm own cumulative incidence.

GPT-6 Astra (gpt-6-astra)

C10-1 (major; Competing risks). Curve increments were converted to hazards incorrectly.

Response (revised). Agreed; verified (hazards 0.4 and 0.2 per year give one year incidences 0.3008 and 0.1504; the previous rule reintegrates to 0.2421 and the corrected inversion recovers 0.4 and 0.2). The error was ours: in round 9 we replaced the derivation fable-5.1 proposed with an incorrect one. The text now uses the exact inversion under constant hazards, H = -log{S(t_e)/S(t_s)}/Delta and h_k = H Delta F_k/(Delta F_1 + Delta F_2), plus the Kaplan-Meier form, as all three reviewers proposed.

C10-2 (minor; Competing risks). Competing mortality curve exception conflicts with first event risk sets.

Response (revised). Agreed; your stricter rule adopted. Curves qualify only as cumulative incidences of mutually exclusive first events, or as Kaplan-Meier curves of each first event with the other censored; a mortality curve that includes deaths after a nonfatal target event cannot substitute. This replaces the dependence assumption fable-5.1 proposed in round 9, which is no longer needed because such curves are not used.

C10-3 (minor; Competing risks). Failure wording overrode a valid subdistribution conversion.

Response (revised). Agreed; when no support source exists, only the exponential approximation is unavailable and every other eligible route, including a reported Aalen-Johansen incidence for subdistribution conversion, is kept.

Fable 5.1 (fable-5.1)

F10-1 (minor; Competing risks; data items). Biased per interval estimator; curve estimator unrecorded.

Response (revised). Agreed; your replacement paragraph adopted, with the estimator of each curve recorded in the data items. Our round 9 response recorded F9-4 as agreed although the derivation was replaced by an incorrect one; this is corrected here.

F10-2 (minor; Competing risks). Reported intervals undefined for curves.

Response (revised). Agreed; intervals are those of reported events and person time or, for curves, between published numbers at risk.

F10-3 (minor; Competing risks). No variance source; exponential route pooling and zero counts.

Response (revised). Agreed; variance 1/d_k from reported or reconstructed counts, unavailability with a zero count or without numbers at risk, and the same pooling and zero count rule for the exponential route.

F10-4 (minor; Absolute effects). Low and high risk scenarios under competing risks.

Response (revised). Agreed; the scenarios use each covering arm own cumulative incidence by the route in use.

Grok 4.7 (grok-4.7)

G10-1 (minor; Competing risks). Curve formula is not a piecewise constant hazard; grid and curve pairs.

Response (revised). Agreed; see C10-1 and F10-1. Both allowed curve pairs are handled (event free survival gives F_2 = 1 - S - F_1), and hazards are estimated on each arm own intervals and carried onto the grid.