Z-axis calibration comparison

Completed: 11 of 12 fits converged. This comparison tests whether the completed control's globe preference survives when each Z-axis gyro correction is removed separately. An IMU (inertial measurement unit) measures turning and acceleration.

Fixing Z gain to zero completed all six fits and preserved globe over flat, then rotating over still globe, under both processing cases. Fixing Z offset to zero completed five of six fits: the still-globe fit with no upstream removal exhausted its evaluation allowance. The other processing case favors globe. The complete offset comparison remains numerically unresolved, and rotation is withheld for it.

All four rotating-globe fits converged, with horizontal error about 0.082–0.085 m and vertical error about 0.696 m, close to the original clock reference. Their costs increased by less than 1%. This control's close rotating-globe fit therefore does not require either Z correction individually to reach its old limit; the remaining free correction may still compensate. The test does not establish the instrument's true calibration. These are repeats of one stretch, not independent flight confirmations. The 12 starts used 1,837 evaluations and about 94 minutes of fitting. No retries are permitted.

The target is the same preselected 7.30-minute stretch of ILVIS0_applanix_57300_610_LVIS_38614965.013, IMU type 21, median ground speed 878.1 km/h. No new flight interval is selected.

The completed clock comparison is the unchanged reference. It favors globe over flat, then rotating over still globe, under both processing cases. Its Z gyro gain and constant offset still reach their permitted limits. The saved calibration audit describes that behavior.

The new alternatives fix Z gain to zero or fix Z offset to zero, leaving the other correction free within its existing limit. Zero is a diagnostic reference, not a measured factory calibration. Each alternative receives rotating globe, still globe and stationary disc fits under both existing processing cases. Other bounds, packet-header timing, original optimizer, initialization and receiver covariance remain unchanged. Interpret shape first and rotation only after consistent globe preference under both processing cases.

The allowance is 12 starts, one per identity, at most 200 evaluations each and no retry reserve. At the previous measured runtime, full allowance use is roughly 2–2.5 hours, plus preparation/publication. One detached worker uses two numerical threads. A durable start journal, exclusive lock, environment/source/input hashes and atomic results support recovery without repeating interrupted starts.

Frozen plan records the full comparison and separate future prediction proposal. This run performs no held-out fitting or archive expansion. Completed studies remain preserved; no statistical decision threshold is introduced.

The numerical allowance is spent. Do not relaunch the worker. To publish the completed evidence without fitting:

env PYTHONPATH=analysis OPENBLAS_NUM_THREADS=2 OMP_NUM_THREADS=2 MKL_NUM_THREADS=2 ~/venv/bin/python analysis/tests/ilvis0_calibration_trial.py --package-only

Never launch a duplicate worker or resume completed numerical work. After completion, --package-only publishes evidence without fitting.