Several areas passed the three review rounds by reading only, or with a single sampled scenario — the round-3 log-rank-split finding (#287) sat right next to a Weibull-deviance split that was passed as "sound on reading", which is the cautionary tale. These deserve a proper numerical pass:
Read-only reviewed (no simulation was run):
- The Weibull-deviance tree split (
surpyval/beta/ml/**): verify the deviance statistic on hand-computed small nodes, split selection on separable Weibull-mixture data, and the warm-start/parent-window monotonicity construction
- Destructive degradation (
surpyval/degradation/**): parameter recovery from simulated destructive samples; direction inference
- Two-stage ADT confidence bounds: coverage simulation for the generated-regressor correction (does the second-stage CI cover the true life-model parameters ~nominally?)
- Recurrent/degradation serialisation depth: round-trip all recurrent and degradation model types (only tree/forest round-trips were verified in round 3)
Reviewed once with sampled depth (soft-clean verdicts):
- Frailty: MC calibration of the theta standard error and CIs (one scenario checked in round 2)
- Fine-Gray under heavy ties and with count weights n > 1 (round 2 verified continuous times only)
- Stratified Cox fitting (the stratified log-rank was verified; the stratified partial-likelihood fit, baseline-per-stratum predictions, and diagnostics-refusal path were not)
Method as in previous rounds: simulate with known parameters, compare against closed forms / R conventions / lifelines / brute force; findings only count with a numerical reproduction.
Several areas passed the three review rounds by reading only, or with a single sampled scenario — the round-3 log-rank-split finding (#287) sat right next to a Weibull-deviance split that was passed as "sound on reading", which is the cautionary tale. These deserve a proper numerical pass:
Read-only reviewed (no simulation was run):
surpyval/beta/ml/**): verify the deviance statistic on hand-computed small nodes, split selection on separable Weibull-mixture data, and the warm-start/parent-window monotonicity constructionsurpyval/degradation/**): parameter recovery from simulated destructive samples; direction inferenceReviewed once with sampled depth (soft-clean verdicts):
Method as in previous rounds: simulate with known parameters, compare against closed forms / R conventions / lifelines / brute force; findings only count with a numerical reproduction.