Completed stopping trial: why the sixth fit remains unfinished

Five of six fits completed. The verified-stopping revision did not qualify. The remaining stationary-disc fit reached a nearly stationary point, but its bounded local stability check failed the declared measurement-step limit. No saved fit is relabeled.

An IMU (inertial measurement unit) measures turning and acceleration. This audit uses saved results from the same 7.30-minute IMU21 control, median ground speed 878.1 km/h. It performs zero optimizer starts and zero native predictions.

Exact stopping evidence

The unfinished fit's stationarity ratio is 9.35e-06 against the unchanged 10⁻⁴ projected-gradient threshold. Its saved final cost and gradient match the previous fixed-scale attempt. The last accepted callback is at evaluation 191; only 1 terminal callback passes both cost and residual-step stability checks. Three are required for the callback exit. SciPy did not report successful termination.

The additional bounded linear correction predicts a relative cost change of 2.4148 × 10⁻¹², far below the 10⁻⁸ cost limit. However, its relative residual-vector step is 1.5541 × 10⁻⁶, or 1.5541 times the 10⁻⁶ step limit. The fresh nonlinear probe was therefore not attempted. A failed linear screen is not evidence that a fresh nonlinear probe would fail or pass.

Cost and residual-vector movement measure different things: individual errors can change while their squared sum barely changes. A tiny gradient or cost change alone cannot replace the declared combined stopping rule. The trial used 736 evaluations, including prediction diagnostics, and about 35.6 minutes. The allowance is spent; no retry is authorized.

All five originally completed fits remain completed. Their largest relative training-cost increase is 2.9139e-06%, below the declared 0.1% agreement limit. Qualification failed because the sixth fit remains unfinished, not because those five worsened.

Calibration and interpretation

The unfinished fit retains 15 parameters at limits and a full-Jacobian condition ratio of 5.15e+07. Its weakest direction mixes gravity_delta (0.901), gyro_gain_y (0.236), gyro_gain_x (0.224), gyro_gain_z (0.184), accel_bias_z (0.180). These normalized loadings describe a local full-Jacobian direction, not feasible confidence limits or a physical calibration measurement.

The completed no-removal prediction comparison still favors globe. The removal-allowed comparison remains unfinished, and rotation is withheld for the combined comparison. An unfinished flat fit does not provide positive flat-Earth evidence. Numerical stopping does not settle the separate clock, calibration or uncertainty questions.

Fit checks, parameter values, and weak directions preserve the saved numerical evidence. The matched results retain track errors and the scientific status.

What to do next

Do not loosen the stability threshold merely because this result misses it by a factor of 1.55. Do not rerun the same failed identity or extend the completed allowance. Prepare independent physical controls for the calibration/timing interaction and alternative numerical strategies. In parallel, audit existing timestamp and residual evidence as specified in the focused diagnostic plan. Joint Z gain–offset profiles and any fresh observed test require their own finite scope. Recording replication and a corpus rerun remain deferred.

Reproduce this saved-data audit without fitting:

env PYTHONPATH=analysis ~/venv/bin/python docs/ilvis0-verified-stop-diagnosis-20261011/build_report.py