Full-pipeline replay: useful partial geometry, weak statistical separation

All six authorized analyses completed and converged. This is progress beyond a design calculation: the pipeline recovered orientation from simulated measurements and retained useful partial model comparisons under both wind treatments. It still does not establish a strong or validated Earth-model result.

The fixed Chicago–Los Angeles window and IMU turns were analyzed under each of the three physical-model truths, with both wind candidates and combined wind plus mixed IMU drift. One shared development seed makes the candidates matched. The saved inputs to the fitter were independently checked identical between candidates for each truth, across all 23 stored input fields. Frozen source, input and numerical-environment identities match.

The run took 335.20 seconds, about 5.6 minutes, using at most two workers with one numerical thread each. The six attempt times sum to 376.53 seconds; that is not the wall time. There were no processing failures or unknown runtimes. All six attempts are now spent.

Each analysis retained 99 cruise minutes, adequate heading diversity and six mount epochs. Reported forward-axis uncertainty ranged from 0.392° to 0.572°, with the provisional matched-reference method. The magnetic watchdog excluded no segments. Real coverage of the orientation uncertainty remains unvalidated. Global model-test rank was one throughout.

The complete three-model gate still excludes every analysis because one contrast is not identified. Pairwise eligibility is conditional on the retained bins and recovered axes. The earlier nominal design screen used supplied axes and a fixed 5° orientation uncertainty; it is a different calculation and failed the broader wind candidate. The narrower recovered uncertainty and retained measurement geometry here must not be substituted into that older planning result or treated as unconditional protocol guarantees.

No bootstrap samples or empirical thresholds were requested in this processing pilot. Thus every pair abstains, including those with usable geometry. There is no calibrated winner, false-rejection bound or power estimate. A pair coordinate is not a three-model decision, especially for a generating truth outside that pair.

Why the planning calculation looked stronger

An audit of the saved analysis outputs found residual-based gyro scales of about 38.3°/hour. The design grid assumed 3 or 6°/hour. These residual scales include simulated sensor effects, processing error and possible model mismatch; they are not measurements of actual WT901 performance. A subsequent saved-data decomposition identifies the dominant cause below.

For the stationary-globe/disc comparison, the fixed-noise design reports information values of about 27–41, whereas the free-fit calculation with observed residual weights reports about 0.52–1.36. The states, weights and conditioning differ, so this is a diagnostic comparison rather than a simple correction factor. The nonlinear profile's standard error is about 0.62–0.64 on a coordinate where the two endpoints are one unit apart. That leaves substantial overlap even though the geometry retains the contrast.

The dominant loss comes from motion correction

The accelerometer measures gravity together with the effect of aircraft acceleration. The analysis currently treats its direction as the gravity direction. When acceleration changes, that apparent direction can move even if the IMU does not tilt. Subtracting its rate of change therefore introduces a false gyro correction.

An independent reconstruction of the prescribed replay attitude, using only saved inputs and parameters, found the following. No flight was generated and no optimizer was run.

The earlier38.3°/hour weighting scale comes from a preliminary fit; it is frozen before the final fit and is a different quantity from the final residual RMS. The reconstructed accelerometer direction agrees with the stored direction within0.0035°, but apparent and true gravity differ by up to2.52°. The prescribed aircraft-roll contribution is only 8.41°/hour RMS and does not explain the dominant correction error. Fitted science/orientation discrepancy is about1.1°/hour. Components are correlated; their squared rates must not be added as independent variances. The full decomposition closes to numerical precision.

True gravity here comes from simulator attitude, which is unavailable in an actual recording. Its substitution is a diagnostic, not an operational fix or new Earth-model result. It also leaves a material motion discrepancy. The next work is an acceleration-aware motion and orientation treatment using observed position and IMU data, including uncertainty from position derivatives, gaps, wind and orientation. It must preserve the frozen replay and must not use simulator truth to make real-flight decisions. Independent physical fixtures already demonstrate that acceleration alone can produce the false tilt correction.

It is premature to launch a large calibration campaign merely because the geometry gate passes. Hardware characterization, realistic uncertainty, justified wind assumptions and separately frozen calibration and independent validation remain necessary.

review.json preserves fit, geometry, orientation, profile and runtime summaries; noise-review.json preserves the saved noise/information comparison; residual-decomposition.json preserves the motion/sensor/science attribution and input hashes; verification.json records the independent matching/provenance checks. The complete six-task evidence is in run/campaign.json, the append-only shard journals and task-linked diagnostic files.