Does Z-axis calibration explain the model separation?
All 27 fits converged. Every fixed gain/offset combination favors globe over disc, then the specified background field over zero field.
An IMU (inertial measurement unit) records turning and acceleration. This audit uses saved fits of one 438-second IMU21 recording,
ILVIS0_applanix_57300_610_LVIS_38614965.013, with median speed 878.06 km/h and minimum 859.55 km/h.
It covers packet-header timing and the no-upstream-removal assumption. These are matched checks, not independent flights.
| Model | Horizontal error range (m) | Objective range | Remaining parameters at limits |
|---|---|---|---|
| sphere_rotating | 0.0805–0.0928 | 2.30818–2.36769 | 0–1 |
| sphere_still | 28.9979–29.5504 | 10748.6–11154.4 | 12–13 |
| flat_still | 63.5504–64.1004 | 49303–50171.9 | 13–13 |
What the grid resolves
The preferred model’s objective changes by only 2.58% across the full grid. Its excellent track fit does not require Z gain −2% or Z offset +1°/hour: zero corrections and every other tested combination preserve the advantage. Those two corrections were fixed, so they are excluded from the remaining-boundary counts above. The preferred model reaches at most one other limit; the alternatives reach 12–13. Boundary counts must therefore be interpreted separately for each model. This result removes an explanation in which the preference depends on the two extreme Z corrections. It does not establish that the measurement model or uncertainty is fully adequate.
What calibration remains uncertain
The best sampled preferred-model point still lies at Z gain −2%, offset +1°/hour, but its advantage over other calibration points is small. The saved fixed-coordinate derivatives and profile costs show this direction of improvement; they do not identify the factory calibration or instrument drift. The weakest remaining local direction at the zero-correction point is dominated by gravity correction and X/Y gyro gains for all three models. The weak-direction table retains every loading and its box-feasibility flags. These describe the saved normalized coordinates, not confidence intervals. No new Jacobian was calculated. Fixed-coordinate gradients are not feasible search directions within each locked fit.
Residuals and what follows
Saved epoch statistics report physical errors, innovation RMS and lag-one correlation. These are in-sample fitted residuals. They do not independently estimate GPS noise or certify a calibrated likelihood. The next matched grid allows a common upstream background-field-removal fraction. A separately predeclared September 24 recording then tests transfer of this robustness. Different dates do not establish independent instruments. Both stages retain the same calibration bounds and finite grid; neither uses Group1 fused navigation as inertial evidence.
Background-field comparisons describe measured-rotation predictions, not their physical cause. This audit adds zero optimizer starts and zero native predictions. It preserves the original 53-stretch findings.
Reproduce
env PYTHONPATH=analysis OPENBLAS_NUM_THREADS=2 OMP_NUM_THREADS=2 MKL_NUM_THREADS=2 ~/venv/bin/python docs/ilvis0-z-profile-audit-20261011/build_report.py