Why the final flat-disc fit is still unfinished
The fixed-scale solver reached a nearly stationary flat-disc fit, but did not complete its required stopping procedure. All five previously completed fits stayed completed with closely matched objectives. The sixth remains unfinished; this audit changes no labels.
An IMU (inertial measurement unit) records turning and acceleration. This audit reads the saved results from the same IMU-type-21 recording used in the prediction trial: 2015-10-05, 7.30 minutes, median ground speed 878.1 km/h. No original measurements are reprocessed. It performs zero optimizer starts and zero native predictions.
What improved and what did not
The unfinished flat-disc fit's stationarity ratio fell from 39.59 times the permitted threshold to 9.35e-06. Its fresh scaled projected gradient was about 9.35 × 10⁻¹⁰, below the unchanged 10⁻⁴ threshold. The training cost changed from 1352.610900 to 1352.283361, a 0.0242% decrease. A small gradient describes local first-order stationarity within the permitted box; it does not prove a global optimum, unique calibration, adequate observations or statistical significance.
SciPy did not report successful termination before the 200-evaluation allowance ended. The last recorded callback was at evaluation 191; its cost matches the fresh final point. It met the measurement-step and cost-change stability thresholds, but only one consecutive recorded callback met both. The earlier step changed the residual vector by about 6.15 × 10⁻⁵, above the 10⁻⁶ stability threshold. Three consecutive stable callbacks were not recorded. The intervening evaluation count does not establish additional accepted steps or stability.
The replay below uses the already published three-step diagnostic thresholds. It is a historical diagnostic, not evidence that a different stopping rule would have terminated these fits at these exact points.
| Processing case | Model | Saved status | Stationarity / threshold: old → new | First three stable callbacks: evaluation | Stable callbacks at end |
|---|---|---|---|---|---|
| bias1_unsubtracted | sphere_rotating | converged | 0.00497 → 0.00548 | 95 | 4 |
| bias1_unsubtracted | sphere_still | converged | 0.00124 → 0.00133 | none | 1 |
| bias1_unsubtracted | flat_still | converged | 8.7e-07 → 9.84e-07 | 141 | 5 |
| bias1_profiled_removal | sphere_rotating | converged | 0.0103 → 0.063 | none | 1 |
| bias1_profiled_removal | sphere_still | converged | 0.000533 → 0.000577 | 96 | 3 |
| bias1_profiled_removal | flat_still | unresolved | 39.6 → 9.35e-06 | none | 1 |
Some successful SciPy stops have fewer than three stable callbacks. Their original convergence criterion was SciPy success plus the fresh stationarity check; the replay neither invalidates those saved results nor promotes the unfinished one.
Calibration and weak directions
| Calibration parameter | New fits at a permitted limit |
|---|---|
| gyro_bias_z | 6/6 |
| gyro_gain_z | 6/6 |
| gravity_delta | 6/6 |
The unfinished fit has 15 parameters at limits. Its full Jacobian condition ratio is about 5.15e+07 in normalized parameter coordinates. 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 are absolute loadings; the overall sign of a singular vector is arbitrary. The full direction is blocked by box limits in both signs. Full-Jacobian conditioning therefore cannot be read as feasible parameter uncertainty. Repeated limit saturation still requires a calibration/observation adequacy explanation. Numerical completion alone cannot settle it.
Fit diagnostics, signed parameter changes, three weakest directions per fit and complete recorded callback histories preserve the details. The original prediction audit documents the Z gain reversal between whole-stretch and training-only fitting.
Next numerical test
The evidence supports testing termination based on scaled stationarity with an independent stability verification, rather than treating SciPy's unscaled stopping flag as the only successful exit. It does not justify gradient-only acceptance, an automatic budget extension, looser physical bounds or relabeling completed studies. See the proposed stopping-rule test. Qualify that implementation on known-optimum and false-stop controls before requesting a separate bounded airborne trial. Do not advance the current fixed-scale candidate to recording replication.
The completed no-removal prediction comparison favors globe. The full comparison remains unresolved because the removal-allowed flat-disc fit is unfinished. That failure is not equal support for flat and globe. Rotation remains withheld by the existing hierarchy.
To reproduce this report from saved evidence:
env PYTHONPATH=analysis ~/venv/bin/python docs/ilvis0-solver-diagnosis-20261011/build_report.py