Can more of the airborne comparisons finish?

Completed: 10/12 fits converged, compared with 11/12 in the baseline. The control still favors globe and then rotation; the diagnostic stretch remains unresolved. This trial did not improve completion. See the audited results and the saved-fit calibration/adequacy audit. The original numerical method and finite scope below are preserved; no further starts are allowed.

This is a small numerical refinement of the completed NASA airborne analysis. An inertial measurement unit (IMU) records acceleration and angular motion. The analysis integrates its original increments and compares the predicted track with the embedded GPS receiver observations under each physical model.

The completed study left 41 of 53 selected stretches unresolved. The saved-result review found that most unfinished fits exhausted their numerical allowance. This trial tests whether explicit numerical stability checks make more comparisons finish. It does not alter the measurements, physical models, nuisance bounds or completed evidence.

Two matched stretches

Both contain IMU21 Group-4 records and have the same recorded installation signature. They are analyzed separately; matching signatures across days do not establish continuous physical installation. Their original intervals remain exactly as selected before Earth-model residuals were examined. Selection of this diagnostic pair uses completion and observable matching, not a desired Earth-model cost difference.

There are 12 approved starts: three models under two processing assumptions for each stretch. Each start permits 200 native evaluations, including initialization, fresh derivative verification and residual diagnostics. There is one start per identity, no retry reserve and no automatic expansion. An interrupted start counts and its missing result remains unresolved; restarting the worker does not retry it.

What changes

The physical prediction, automatic derivatives, receiver covariance, nominal 200 Hz clock, calibration envelopes and initialization stay the same. Initialization includes the baseline's force-derived accelerometer gains and zero Earth-rate removal.

The new solver checks three consecutive accepted steps for all of:

• Scaled projected gradient at most 1e-4, preserving the independent stationarity threshold.
• Relative cost change at most 1e-8.
• Relative change in the predicted, weighted measurement residuals at most 1e-6.

The raw gradient fallback is 1e-14; the parameter-step tolerance remains 1e-11. A raw numerical stop is accepted only with independently verified stationarity and measurement stability. If it stops before stable steps are available, a bounded linearized correction is checked with an actual native evaluation, within the same allowance. A correction that materially changes the predicted observations fails that check. Budget exhaustion never establishes convergence.

The final projected-gradient components, bound signs and Jacobian singular vectors are saved. They can help identify weak combinations of nuisance parameters. This does not remove those directions or invent a new rank cutoff. Stable track prediction does not mean every calibration parameter has been uniquely determined.

Analytic tests cover interior and boundary optima, redundant and poorly conditioned parameters, a nonlinear known optimum, budget exhaustion and invalid observations. The source, inputs, environment, method and identities are frozen before observed fits.

Interpretation and continuation

Compare finished fits, costs, residual consistency and parameter limits against the saved baseline. Interpret globe/disc first; report rotation only after the required shape comparisons consistently favor globe. Numerical completion is not model support. The report will state which model fits better and whether more calculations finish. It does not assign statistical significance or certify calibration, timing, onboard corrections or receiver errors. The separate scientific significance gate remains unchanged.

Runtime state is in data/ilvis0-solver-trial-20261010/status.json; charged starts and hashed per-fit results are saved there. The worker publishes a matched report and full receipts here on completion without refitting. New costs and failures are preserved whether or not discrimination improves. A changed control result or materially worse objective requires investigation before considering a larger trial.

Use at most two numerical threads. To launch or resume unfinished, unchanged work:

env PYTHONPATH=analysis OPENBLAS_NUM_THREADS=2 OMP_NUM_THREADS=2 MKL_NUM_THREADS=2 ~/venv/bin/python analysis/tests/ilvis0_solver_trial.py --detach

The exclusive worker lock is inherited by the detached process, preventing duplicate numerical workers. Completion prohibits further starts. After completion, use --package-only to regenerate the audited report without any model evaluations.