Local Level Lab

Airborne measurements. Open methods.

Horizontal-only recorded-flight development study

Question and selection

Does the globe preference survive when the models are refitted using horizontal receiver positions, with receiver height removed from the scored objective?

Use the shortest stretch with all six reusable primary artifacts for each of IMU8 and IMU21. Choose by duration, filename, start time and segment ID, in that order, from the completed artifact inventory and original frozen selection. Neither full-position nor projected model costs enter selection. This gives two complete stretches of about 585 and 244 seconds on distinct dates.

Fit the complete saved intervals. Do not recalculate eligibility, shorten the intervals or select additional stretches after inspecting these results.

Matched physical assumptions

Retain all raw angle/velocity increments, the nominal 200 Hz integration clock, receiver epochs, initialization and h1_v3_tau60 covariance assumptions of the existing study. Marginalize receiver covariance before whitening; score only latitude and longitude. Keep all 26/27 nuisance columns, original bounds and both fixed-zero/profiled upstream-removal cases. Constant gyro offsets remain bounded at ±1 degree/hour, an assumption rather than a measured unit limit.

Use the newly qualified ridge-probe candidate and extended tangent arithmetic. Its displacement penalty changes only the probe, never the physical-model score. Final stationarity, cost and physical-prediction limits remain unchanged.

Receiver heights remain in initialization, covariance construction and trajectory propagation. Both gyro and acceleration measurements remain. This is a horizontal-objective sensitivity study, not altitude or datum independence and not a gyro-only test. Group1 fused attitude/rates are not fitting observations. The existing full-position fits are historical references, not matched reruns under this new numerical method. No percentage of evidence is assigned to a channel.

Fixed identities and allowance

Two stretches × three models × two processing cases × two starts = 24 starts. The primary starts at the exact current saved full-position parameter vector; its repeat starts from the original receiver/IMU initialization rule, independently of fitted Earth-model residuals. The same initialization code applies to each model.

All baseline systems must reproduce full costs and residuals to 1e-6 relative scale and Jacobians to 1e-5, with a unit floor in norm comparisons. The tolerances allow the documented ordinary/extended arithmetic difference; report actual differences rather than calling them bit-identical.

Acceptance and reporting

Every model/case primary and independent repeat must pass the unchanged fresh local gate. Require paired cost agreement within 1e-8 relative to max(costs,1), axis prediction RMS within 1 mm and epoch maximum within 1 cm. These are numerical limits, not sensor precision or calibrated scientific uncertainty.

Compare globe family against stationary flat disc first, consistently across both processing cases. Report the field comparison only after completed shape comparisons favor globe. An unfinished fit or failed repeat withholds the matched comparison; preserve its costs and failure. Keep the existing 53-stretch tally.

Save the full returned parameter points, projected residuals/Jacobians, actual receiver times, coordinates, unscored height errors, per-axis residual summaries, start/call journals and source hashes. No production promotion or significance claim follows. Publication verifies saved artifacts and never triggers refitting.

Implementation and checks

Implement and verify the metadata helpers first, then the worker. Launch only after the passed software completion and all frozen inputs have been verified. Preserve all earlier completed studies; no unused allowance is transferable.