Skip to content

A second review, of the fixes

A fresh context reviewed the previous day's corrections without being told why they were written. Nothing blocking; every committed result it regenerated reproduces from HEAD, and it confirmed the factor geometry computes A minus the broadcast initial A and records that file's hash. Its serious finding was about the tests rather than the code: 24 of 59 mutants survived the suite, among them the probe fix's own regression test, which could not fail on a mutant that restored the very behaviour it was written for, because its fixture scored 1.0 and only a label on the result told the versions apart.

Fixed here:

  • The probe test now counts fits. One fit per fold, whatever the number of resamples, which the restored-refit mutant fails. The fixture now scores about 0.55 so an interval has room on both sides, and a second test holds the estimate and the interval to the same predictions.
  • accuracy_interval cross-validates once. It had fitted every fold twice, once for the estimate and once for the resampling rows.
  • A rank that spans the held-out clients is no longer scored. The bases are fitted without a whole project, so the largest project sets the limit; past it the shared half is the whole space and the residual is numerical noise (largest entry 4e-15 on 34 clients at rank 32) that normalises into unit vectors and scores like an ordinary cell. The committed result is unaffected, since its smallest held-out fit holds 37 rows, but the registered test on the ten-project rebuild would have met it at the default ranks.
  • A third organization is refused by the subspace null instead of being folded into the second, which would have put the truth outside its own null and cost the p-value its floor.
  • energy_in_shared is renamed energy_in_global_shared, since it describes the global basis the detector uses and not the held-out bases the attribution uses.

Still open from that review, and recorded rather than fixed: a SIGKILLed job leaves a zero-byte claim that cannot be told from a live one, which a Slurm requeue would refuse; several load-bearing behaviours have no test (the permutation's split over projects, the family-wise step, the contamination partial's claimed name, calibration claiming every condition before training).