Fitting wind and aircraft-motion correction together

The acceleration correction must depend on the same wind and orientation assumptions as the Earth-model prediction. Otherwise the fit could adjust aircraft heading while its motion subtraction stayed fixed at a different heading.

The joint research prototype now recomputes both at every trial parameter value. Observed GPS velocity and IMU force determine acceleration and apparent vertical direction. The wind candidate determines aircraft heading. The uncertain forward reference determines how that heading maps into the IMU axes. The prototype recovers gravity direction, computes aircraft rotation and the model prediction in the same orientation, then averages them over the same minutes. Its derivatives include both paths through wind and orientation.

Six explicit error parameters represent coherent GPS-acceleration and accelerometer-force errors. They enter the same prediction, with provisional uncertainty scales and priors. The joint covariance therefore includes their coupling to wind, orientation and model coefficients. These six patterns are not a complete model of measurement errors; validated coverage, correlations and instrument systematics remain open.

Time support is fixed before optimization using observable data and a bound covering simultaneous error perturbations. The fit cannot improve its objective by dropping different minutes at different parameter values. All six saved cases retain the same83minutes.

Checks and saved-data work

Complete: all24optimizer calls converged and the six free fits nested below the fixed-model fits. No nesting restart was needed and no fit approached a parameter bound. All cases used the same83minutes. Residual variation after joint refitting is7.92–8.57°/hour: 8.49–8.57for broader wind and7.92–8.03for physical wind/TAS. Completed optimizer time sums to309.86seconds; that is not a measured wall time. Sixteen focused checks pass. A resume check skipped all completed fits, preserving exactly24starts and24completions.

This confirms that the correction survives joint optimization in these saved development recordings. It does not establish calibrated model separation: the weights and error model remain provisional, and no empirical decision is enabled. Matching gyro and GPS/force filter bandwidth and validating measurement covariance are the next processing checks.

Six saved reference states were evaluated without fitting. Both wind candidates passed independent derivative checks. At those earlier parameter values, residual variation was about11°/hour for the broader wind candidate and15°/hour for physical wind/TAS. The older parameters came from a different motion treatment; these are reference checks, not optima. Local coefficient uncertainty remains substantial under the frozen older gyro weights.

A bounded diagnostic refit then uses these same six recordings. It makes one free fit and three fixed-model fits per case, with one predefined free restart only if the objectives fail to nest. The scope allows at most30optimizer calls, including interrupted starts, and200evaluations per fit. No new simulated recording or bootstrap is included. Empirical decisions and production use are disabled.

The runner freezes source, input, policy and numerical-environment identities. One numerical thread is used. Every optimizer start is journaled and fsynced; completed fits have atomic metadata and hash-checked parameter archives. Resume skips those fits, retains interrupted starts and refuses source/environment changes. A process lock excludes duplicate workers. If the call limit is exhausted, the runner stops rather than extending its scope.

Evidence and continuation

This finite scope is complete; no worker is active and no resume is needed. The unused optimizer-call reserve is not permission for additional fits. For interruption recovery in this frozen scope, read refits/status.json before any resume. An active worker needs no duplicate. The command below resumes only the frozen missing work and preserves the30-call maximum:

env OPENBLAS_NUM_THREADS=1 OMP_NUM_THREADS=1 MKL_NUM_THREADS=1 /home/sweisman/venv/bin/python analysis/tests/refit_joint_acceleration_motion.py docs/observed-pair-replay-20261007/run --output docs/joint-acceleration-motion-20261007/refits

Implementation: analysis/tests/joint_acceleration_motion.py; reference evaluator: review_joint_acceleration_motion.py; refitter: refit_joint_acceleration_motion.py; fixtures: test_joint_acceleration_motion.py. All remain outside production analysis.

The gyro weights still come from the original preliminary fits. GPS/TAS speed remains conditionally modeled at bin resolution. Coriolis/curvature acceleration bounds, complete measurement covariance, filter bandwidth and independent uncertainty validation remain necessary. No projection diagnostic here changes the frozen eligibility gate, and no raw objective difference is a calibrated model rejection or an Earth-model conclusion.