Why calibration stopped
Both starts of one required setting failed the numerical cost-stability check. The phase therefore stopped under its predeclared rule; validation did not start.
This review reads eight records selected before inspecting their fit details: seven saved fits and one interrupted start. The failed pair is compared with the preceding noise example and the available alternate-processing fit. No observations were generated, no fits were rerun and no native calculations occurred.
| Saved check | Failed pair | Required limit |
|---|---|---|
| Evaluations used | 41 each | At most 200 |
| Fresh stationarity | About 0.0000356 | At most 0.0001 |
| Largest fresh-check position RMS change | About 0.225 mm | At most 1 mm |
| Largest fresh-check position change | About 0.482 mm | At most 10 mm |
| Relative cost change | About 0.000000827 | At most 0.00000001 |
The cost change is about 83 times its frozen tolerance. The predicted positions pass their physical-error screens, but that does not override numerical qualification. Both failed starts report “no positive predicted decrease.” The preceding example passes all four selected starts; the available alternate processing case also passes. The interrupted flat-model start has no completed fit and is not treated as a numerical or physical model failure.
The reusable saved-record reader and reporting adapter passed seven software
checks. The immutable evidence snapshot preserves the request, exact selected
records, their hashes, reporting source and unchanged failed labels. Reproduction
uses lll.prediction_phase_review.run and the frozen request in a separate
working copy; the original one-review allowance is spent.
Selected records CSV · Review plan · Phase outcome · Calibration evidence.