Recovery stopping diagnosis

The trial is complete: 15 of 24 fits passed, including all six completed-control fits and eight of the 15 originally unfinished primary fits. The three unresolved stretches still lack complete comparisons. The original public findings remain 12 globe preferences, zero flat preferences, 41 unresolved.

This audit reads sealed results and journals only. It makes no new predictions or fits, and changes no scientific labels. The 1,694 charged sweeps are well below the 14,448 ceiling; the unused allowance cannot be reused.

Why the nine failures differ

Four early stability failures: the saved gradient passed, but a bounded linear correction did not meet the unchanged measurement-stability requirement.
Four early stationarity failures: the optimizer stopped on parameter-step size while the independently checked gradient remained excessive.
One budget failure: the perturbed disc fit remained far from stationary when its fitting allowance was spent.

All eight early exits used the optimizer's xtol parameter-step test. Small parameter steps are therefore insufficient evidence of completion. None of the four failed stability probes reached a fresh nonlinear correction measurement: they were rejected by the linear prediction first. Whether those corrections improve the actual observations is not saved and cannot be inferred here.

Stretch prefix	Processing case	Model / role	Evaluations	Gradient / limit	Probe result
03454752	profiled_removal	flat_still / repeat	598	350.8	not_stationary
03454752	profiled_removal	sphere_rotating / primary	32	1.627	not_stationary
233183f6	profiled_removal	sphere_rotating / primary	18	0.0002222	linear_correction_not_stable
233183f6	unsubtracted	sphere_rotating / primary	19	0.0001184	linear_correction_not_stable
233183f6	unsubtracted	sphere_rotating / repeat	29	3.788	not_stationary
b365bdf3	profiled_removal	flat_still / primary	94	0.04737	linear_correction_not_stable
b365bdf3	profiled_removal	sphere_rotating / primary	157	1.072	not_stationary
b365bdf3	profiled_removal	sphere_still / primary	24	0.8403	linear_correction_not_stable
b365bdf3	unsubtracted	sphere_rotating / primary	90	1.02	not_stationary

A gradient/limit ratio below one passes stationarity. It does not establish measurement stability. Three of the four failed stability probes predict tiny cost changes while predicting residual changes above the allowed limit; the fourth predicts a larger cost improvement as well.

The failed Jacobians have a median condition ratio of 1.13e+08. That measures the disparity between strong and weak local parameter combinations; it is a warning about numerical sensitivity, not proof of its cause. Saved right singular vectors and complete Jacobians are absent here, so the weak combinations cannot be reconstructed. The dominant remaining gradients include gain/offset/gravity combinations and near-boundary calibration coordinates.

One accepted fit passed three stable callbacks although its linear probe failed. That is permitted by the frozen rule, but shows that callback stability and full linear-correction stability are different tests. It must remain visible in the next qualification; this audit does not change either tolerance to improve counts.

Next numerical test

Screen a genuinely different bounded step strategy on deterministic equations with known optima, coupled weak directions and near-boundary starts. Preserve the same gradient, cost and measurement-stability requirements. Compare objective and predicted observations, including failed tests. Do not select a strategy by its preferred Earth model. An observed diagnostic must then record rejected steps, fresh correction behavior and local tangents under a separately frozen finite allowance. No new observed fit or all-41 rollout is authorized by this audit.

Reproduce

Run from the repository root:

env PYTHONPATH=analysis ~/venv/bin/python docs/ilvis0-recovery-diagnosis-20261012/build_report.py

The builder checks the original completion seal, manifest, summary and both charged journals. fits.csv preserves every accepted and failed result; the receipt hashes inputs, audit software and generated artifacts. Original data and completed trial outputs remain untouched.