What remains in the development plan

The main implementation work from the review is present. The remaining task is demonstrating that useful flight geometry survives realistic instrument and aircraft uncertainties, then calibrating and independently validating the resulting decisions. Implemented software and successful numerical fits do not establish that the physical models can be distinguished.

The current development priority is the archival airborne analysis, with the passenger experiment to become secondary once the archival analysis is ready. Six prepared IMU6 recordings now have GPS-only reconstructed trajectories and correlated-error sensitivity checks. Neither is an Earth-model result. Joint IMU/GPS covariance controls now distinguish noiseless calibration/timing recovery from useful precision and preserve GPS-offset/drift ambiguities. Longer rotation controls now test duration and brief pitch motion; all22checks pass, while joint information remains weak or rank-sensitive. The instrument evidence matrix specifies the missing legacy processing, calibration, mounting and integration-clock support. Information stability diagnostics now compare four central derivative steps and smooth information relative to the declared fixture bounds. This separates numerical rank from useful calibration; the bounds and reference penalty remain assumptions. The instrument documentation review now supports the LN-200-family inference and packages firmware-specific unresolved questions. It does not establish the recorded corrections or calibration bounds. The legacy-record search records the public routes checked and an unsent NSIDC request. The guide confirms raw archive provenance; matched onboard-processing/calibration records were not obtained. The inquiry has been sent. Conditional recorded-data modeling now proceeds on six files with explicit assumptions while a response is pending. It uses the existing24-start/200-evaluation ceiling; the scientific evidence gate remains unchanged. Preserve the six outcomes, then extend to supported recordings in the kept corpus with a finite, separately frozen scope. Existing synthetic studies stay frozen; they are not rerun as part of this archival work.

The six-file pass is now complete:24finite fits,0converged within that allowance. The 232-file continuation is launched under a separate764-start aggregate/200-evaluation scope. It preserves unsupported configuration and convergence outcomes, and verifies fresh derivative stationarity. One interrupted writer attempt stays charged in the repaired v2 freeze. No winner or calibrated scientific gate is promoted.

The next refinement uses instrument-informed offsets and compares shape before rotation. The six-file refinement now implements exact forward derivatives, column scaling and staged initialization with all final nuisance parameters released. Four cases per recording combine ±0.1/±1 degree/hour constant gyro-offset hypotheses with fixed-zero/profiled common Earth-rate removal. Shape must be consistently resolved before rotation diagnostics appear. It queues after the current corpus and uses a separate72-primary/84-total-start freeze,200evaluations per start; no synthetic rerun or scientific promotion. The old ±20 degree/hour allowance was never measured instrument drift.

The broad 232-file run is complete; the six-file refinement is running. The next all-stretch extension saves every supported interval of at least 240 seconds with receiver ground speed at least 700 km/h. A public catalog records exact intervals, criteria and exclusions before fitting, with fingerprints for reuse. Whole stretches are primary joint fits; smaller sections check consistency at the same parameters. The extension waits for the six-file completion and uses a separate inventory-derived finite freeze. Scientific eligibility/calibration remains unchanged.

Recent processing work also addresses nuisance-rank scaling, forward-reference noise/motion handling and finite-run course unwrapping across gaps. Actual forward-reference uncertainty and its statistical coverage still need evidence; software fixtures do not establish it.

For the passenger experiment, the remaining order is:

1. Continue inexpensive route/window/turn screening under the unchanged nuisance space. Evaluate rotation and globe/disc questions separately. Seek an actual observed route, not only a counterfactual direction control.
2. For promising controls, review the finite envelope assumptions and obtain a separately bounded budget for complete-envelope evaluation and matched full-pipeline replay. Preserve every failure and abstention. Do not rerun spent attempts to improve acceptance.
3. Decode the supplied airborne IMU using a documented physical format, and qualify actual devices when available. These are separate evidence sources: the NASA instrument cannot establish consumer-device drift or accuracy.
4. Once a useful domain is demonstrated, freeze source, configuration, eligibility policy, observable domain and numerical environment. Preregister primary, pairwise, pooled and composite-shape error budgets where applicable.
5. Run fresh calibration, freeze its actual decision artifacts, then run independent validation and assess the simultaneous exact bounds. Production use follows a reviewed policy revision, not a development preference.

All previously authorized simulation attempts have been spent; no worker is being launched by this continuation. Earlier calibration costs and prospective execution freezes are stale. No usable fully validated protocol or empirical Earth-model result has yet been demonstrated.

Latest inexpensive check

A fixed reversed traversal of the preserved 120-minute AUH–ORD path, with six IMU turns, previously passed a nominal globe/disc design check under the physical wind candidate. The bounded follow-up evaluated 30 states from the existing registered envelope in about 0.10 seconds. These include zero wind and some individual wind drifts, all three model anchors for completed state groups, and forward-angle offsets of 0° and ±15°. Both globe/disc retained fractions remained above the threshold in that subset: minima 0.403480 and 0.431011 against 0.312250. The evaluation cap stopped partway through the grid. This found no counterexample; it did not complete or certify the envelope. Other wind corners, airspeed levels, mount epochs and noise states remain unchecked here.

The broader wind candidate already fails the nominal check on this same reversed path. The reversed traversal is not a measured eastbound flight, and no sensor synthesis, fitting, bootstrap or calibrated decision was performed. Thus this remains a design clue, not a protocol recommendation. The recorded states and input/helper hashes are preserved in direction-envelope-counterexample.json, produced by analysis/tests/check_direction_counterexamples.py.

The airborne sample notes and format follow-up describe the separate real-data step.

Whole-route two-hour screen

The earlier longer-window search extended only anchors that already had promising 75-minute geometry. A separate check now samples two-hour windows every 15 minutes across all seven recorded tracks, regardless of those earlier outcomes. It uses three fixed patterns: no IMU turns; turns at minutes 25, 50 and 85; and turns at minutes 10, 30, 50, 70, 90 and 110. Coverage, aircraft-motion buffers, duration and heading rules are unchanged.

Of 176 overlapping windows, 130 failed observed coverage and one lacked the required height information. The remaining 45 windows received motion/heading screening. Seven window-and- pattern combinations passed: six on Abu Dhabi–Chicago and one on Chicago–Los Angeles. No eastbound combination passed the heading requirement on this grid. None of the seven passing controls preserved any model contrast above the retention threshold under either wind candidate, even at nominal nuisance settings. These SVD checks use all three model anchors, dynamic sensor bias, forward-angle uncertainty, zero wind/angle, nominal true airspeed of 250 m/s and a fixed 6°/hour noise assumption.

The complete check took 1.80 seconds and used no simulated flight attempts. Six existing motion-screen and direction-control fixtures passed in 0.45 seconds. Results and input/helper hashes are preserved in long-route-window-screen.json; the reproducible helper is analysis/tests/screen_long_route_windows.py. This coarse search is not exhaustive and does not establish impossibility. Other starts, lengths and safe turn timings remain open. No expensive replay, full-envelope calculation, calibration or protocol promotion followed.

The continuation checked three-/four-hour windows independently of the earlier anchors. The longer-route review records 277 additional windows, 250 coverage failures, 27 motion-screened windows and ten preliminary passing combinations. None passed any nominal model contrast; no four-hour combination passed the preliminary screen. Six patterns per duration include shifted central six-turn timings, with all existing schedule and maneuver rules preserved. Runtime was 2.65 seconds; six existing motion/direction checks passed in 0.44 seconds. No flights were simulated. The helper is analysis/tests/screen_extended_route_controls.py; its output is extended-route-controls.json. The invalid initial timing diagnostic is preserved separately and excluded from scientific conclusions. These results do not exhaust possible schedules.

Safe timing search and partial evidence

The turn-timing review records the next bounded search: 88 safe schedules on the 11 previously preliminary-passing windows, with no passes across both wind candidates. Sixteen individual schedule/pair comparisons passed nominally under the physical wind candidate alone. Small registered-state checks disproved five; eleven remain unresolved. No full-envelope claim, noisy replay or empirical decision follows.

The strongest surviving stationary-globe/disc partial control is frozen in observed-pair-envelope-plan.json: actual ORD–LAX elapsed seconds3600–10800, turns5/15/30/40/90minutes and the physical wind candidate. Its complete registered envelope was authorized and completed on 2026-10-07: all15,210states in29.70seconds, one numerical thread, under the120-second cap. The direct-pair retention is at least0.365100 against 0.312250; this partial comparison passes the finite grid. The global rank is1 throughout. Rotation passes narrowly in the direct-pair calculation but fails after the global rank cutoff; rotating globe versus disc fails under either calculation. The completed result is observed-pair-envelope-result.json. No flight was simulated and no threshold calibrated.

Next: prepare and separately budget matched full-pipeline replay of this fixed control across the three truths, combined wind/sensor drift and both wind candidates. Supplied design axes must be replaced by actually recovered axes. Broad wind already fails nominally, so the physical candidate's assumptions need scientific justification even if replay succeeds. The existing flight-attempt allowances remain spent; this one-envelope compute approval is also complete. Do not rerun it or start another envelope/replay without the required budget.

Completed matched full-pipeline replay — 2026-10-07

The matched replay covered six development attempts: three truths × two wind candidates, shared fresh seed600910 and combined wind/bias_mixed. The frozen control is unchanged. All six analyses completed and converged in335.20wallseconds, with 99cruise minutes and adequate headings. Orientation was actually recovered from simulated readings. The two candidates received independently verified identical fitter inputs within each truth. Both rotation and stationary-globe/disc pairwise gates pass in all six cases; rotating-globe/ disc fails, and the complete three-model gate remains excluded. All decisions abstain: zero bootstrap, no empirical thresholds, no calibrated winner or error-rate claim.

The replay summary and its frozen evidence record the new limit: residual-based gyro scales are about38.3°/hour rather than the3–6°/hour planning assumption. Stationary-globe/disc profile standard errors are0.62–0.64 on endpoints one unit apart. The saved-data decomposition is now complete: aircraft-motion mismatch is 39.63°/hour RMS, sensor/calibration remainder3.86–3.93°/hour, and the false tilt correction from treating specific force as gravity is38.12°/hour. Prescribed true-plumb substitution, with saved fit parameters unchanged, leaves about12°/hour. These correlated components cannot be added as independent variances. No fit, new flight or decision was produced. Next develop an observed-data acceleration-aware orientation/motion correction and test its uncertainty before a bootstrap/power pilot or large calibration. True simulator attitude must stay diagnostic-only. Do not attribute these scales to real-device noise. The six attempts are spent; there are no active workers or authorization for additional flights, calibration or promotion.

The observed-data acceleration prototype is now implemented outside production. Independent motion fixtures and saved-data checks pass (11focused tests). On the same85supported minutes, residualRMS falls33.75–33.89to11.12–11.20dph with saved parameters held fixed. All93sensitivity states complete;83minutes have common support. Their residual range9.6–37.3dph shows that wind/orientation coupling remains material. GPS/IMU uncertainty propagation is provisional and assumes independent errors. No new flight, optimizer, eligibility change or decision was produced. Next couple observed acceleration and orientation to the research wind/forward nuisance parameters and verify their joint uncertainty and remaining motion error before any separately scoped refit/calibration.

The coupled prediction and bounded saved-data refits are now complete. Corpus: joint-motion study. The correction varies with the fitted wind/forward parameters and six coherent error modes; its Jacobian/covariance include those paths. Fixed support83minutes. All24optimizations converge and nest; no boundary flags or repairs. ResidualRMS7.92–8.57dph, summed optimizer309.86s. Sixteen focused checks pass; checkpoint hashes and a zero-extra-call resume are verified. No new flight, bootstrap, calibration, eligibility change or model decision. Weights/error priors remain provisional. The saved-data filtering/covariance check is now complete: identical additional filtering of observation and full prediction leaves 79common minutes, reducing residualRMS7.85–8.32to7.08–7.48dph. Slow-prediction interpolation changes only0.026–0.040dph. An explicit gyro covariance propagates filter overlap and axis rotations; independent-second marginal scale~3.53dph is provisional and includes motion. Residual adjacent-minute correlation reaches−0.61; fitted-residual correlation does not identify a noise source or prove significance. All six cases and checks are preserved; 21focused tests pass,0optimizercalls/newflights,0activeworkers.

The continuous-equation/shared-error check is now complete. All terms are evaluated at GPS timestamps; no slow-term interpolation. Six saved free-fit states,79common minutes, residualRMS7.0763–7.4782dph. Shared GPS/paired IMU covariance gives fixed-parameter sigma11.2533–11.2722dph under independent-second assumptions, about90% of variance from GPS. Local old-weight penalized fit response preserves GPS/physical-speed-constraint cross covariance; residualsigma11.1343–11.1603dph. This is conditional sampling propagation at saved states, not a fit or validated coverage. Eight new tests,29focused total; independent full-input derivative and least-squares response checks pass. Six checkpointed evaluations,0optimizer/newflight,0activeworkers; completed-scope resume skips all6.

The temporal covariance study is complete:6saved cases×36GPS/IMU persistence combinations,216scenarios. Durations0/1/5/15/ 60/300s, exponential elapsed-time kernel, fixed marginal scales/79minutes. Sensor-axis IMU roots preserve mount rotations and shared errors. Predicted fixed-state sigma11.2533– 42.3085dph; local fit-response residualsigma11.1343–40.6480dph. Worst residual scale uses GPS15s/IMU300s in every case. This grid is assumed sensitivity, not a credible-domain bound or characterized physical covariance. Baseline recovery, dense-kernel, PSD and all216 response summaries verified;34focused tests pass,67.803summed seconds,0optimizer/newflight/ activeworkers. Resume skipsall6. Next implement/verify a covariance-aware research objective with shared GPS auxiliary errors before separately defining any further finite refit scope.

The frozen-covariance objective study is now complete:216saved-state evaluations, full joint whitening, matching gyro/auxiliary signs, fixed matrices, log determinant and deterministic penalties. Singular matrices fail without jitter/clipping. A second216comparison preserves the existing2m/s physical wind discrepancy as explicit auxiliary covariance. Measurement-only physical-wind quadratics27,718–598,005 shrink to167–336with that registered allowance; broad-wind controls unchanged. These old-state objectives are not calibrated fit statistics. 44focused tests, independent matrix/directional/response checks pass;54.164+7.093summed seconds,0optimizer/newflight/activeworkers; completed resumes skip allcases.

The saved-recording GLS comparison was separately authorized on October 7 and is complete. Six saved cases×4fixed correlation pairs× (free+3fixed)=96primaryfits,≤120starts including predefined single nesting repairs and interruptions, max_nfev200. No bootstrap/newflights/decisions. A source/input/environment freeze, synced per-fit archives and append-only journal support safe continuation. Six cases share three simulated recordings, not six independent flights. All96fits converge; all24comparisons nest;0repairs/interruption/failures/boundary or1%-near-bound fits. Optimizer1732.595s; about30.5minutes launch-to-completion, one numerical thread, no active worker. Free residualRMS4.4711–9.8778dph. Independent residual/objective/ full-equation direction checks, covariance/support identity and start/fit checkpoint audits pass;17focused checks. Resume skips all24comparisons with no additional starts. Injected model has the lowest fixed objective for every assumption/candidate, but differences shrink substantially with longer error persistence. Local data-only contrast retention is23.5–52.4%rotation/still,4.0–14.3%rotation/disc,31.7–64.3%still/disc. These free-state diagnostics apply no gate or calibrated threshold. Weak rotation/disc geometry remains; covariance persistence/reference/calibration coverage remains unmeasured. Next characterize hardware/GPS persistence and validate the motion correction, then define a defensible research-candidate covariance domain and informative protocol. Fresh calibration and independent validation follow; no further campaign is authorized.

Preparing the next measurements

The input-persistence diagnostic is implemented in analysis/tests/characterize_input_persistence.py, with nine controlled tests. It processes explicit still phases separately, retaining incomplete seconds/gaps, IMU/GPS/shared-channel covariance, finite-lag moments, averaging variation and two detrending views. No automatic covariance/threshold selection or qualification. No real-device recording has been characterized; stationary GPS does not establish airborne error behavior. This is software preparation for the next measurements, not a new flight/refit/bootstrap/calibration campaign. The completed source/corpora remain frozen and production science is unchanged.