A small airline-route pilot

An IMU (inertial measurement unit) measures acceleration and turning. This pilot would test the analysis on simulated IMU readings along an archived Frankfurt–Johannesburg position track. It compares three corrections using the same generating conditions: the existing wind correction, the physical wind/airspeed correction, and the physical correction with the optional mount-movement comparison. The last path still requires agreement with the original exclusion method before reporting useful evidence.

The initial three evaluations are complete; all failed before fitting. They used the same simulated flight under three corrections. The IMU/GPS relationship was too weak to establish the aircraft's forward direction. The remaining 15 evaluations have not run and are not authorized. No model separation, calibration or validation was obtained.

Initial result

The forward-reference check compares IMU roll with the bank change inferred from GPS course and speed. All three cases had the same regression gain of 0.0487 and explained fraction of 0.0123, below the existing requirements of 0.1 and 0.05. The check reported weak roll/bank correlation, rather than an absence of turn activity. No orientation epoch was unresolved; 57 bins remained for a possible fit. Passing the buffered hand-turn check did not ensure a usable forward reference.

The existing wind candidate required a measured forward-angle uncertainty; both physical wind candidates required a three-axis forward reference. None reached the science fit, envelope or pair profiles. Their 8.14, 8.25 and 6.39 seconds therefore measure failed preprocessing attempts, not successful-fit cost. Known attempt time totals 22.77 seconds; the saved run occupies 14,958,617 bytes, including three raw simulated sessions. Two single-thread workers completed the approved prefix and stopped automatically.

The next useful work is to examine the saved forward-reference calculation and the assumed trajectory's roll/GPS relationship before proposing another replay. Keep these failures; do not weaken the reference checks, insert the simulator's known forward axis, or retry these spent attempts. This result neither tests the airspeed correction nor establishes failure on an actual flight with its own GPS and IMU measurements.

What the saved-data investigation established

The bank estimate differentiates GPS course twice. Small errors in course can therefore overwhelm the roll signal. A diagnostic using the assumed path without GPS noise explains about 28% of the scalar roll variation, compared with about 1% using the saved noisy GPS. This comparison isolates a numerical effect; it does not reconstruct measured aircraft motion.

A research-only calculation applies the same 30-second smoothing to both the GPS bank-rate estimate and the IMU roll readings. It uses complete supported windows, splitting at missing data or irregular intervals. On these saved sessions it explains about 90% of the horizontal IMU variation, with gain 0.925, without lowering the existing correlation or gain requirements. Independent controlled turns recover the known forward direction; GPS noise without roll does not create a reference. The calculation deliberately supplies no angle uncertainty or scientific acceptance. Uncertainty must account for filtered, correlated data and errors in the GPS predictor before this can become a fitting method. One development result is not evidence of calibrated coverage.

The original position interpolation has a second problem: its acceleration can change abruptly at observation times. Inferring coordinated bank from that acceleration creates instant bank steps, up to 15.9° here. At the simulated 20 Hz sample rate the assumed scalar roll reaches about 162°/s. About 93% of scalar roll energy falls within a second of an interpolation knot. These are artifacts of the assumed path, not observed aircraft turns.

An explicit alternative uses a cubic curve with continuous acceleration inside each supported block. It preserves every supplied position and height, and the same observation gaps. On this window the largest scalar roll rate is below 0.34°/s; the finite one-sided knot check is below 0.00031°. This removes the bank-step artifact, but its endpoint curvature and motion between fixes are still assumptions. It may overshoot on other tracks and must be reviewed. The original interpolation remains the default for historical specifications; the alternative has its own version and trajectory hash. No flight has been generated from it.

The next implementation step is an uncertainty-aware matched forward-reference method, with controlled tests of GPS error, missing data and real turns. Then a separately budgeted, freshly frozen pilot can compare the smoother route through the full pipeline. Do not resume the old plan after scientific-source changes or use these diagnostics as replacement winners.

That method is subsequently implemented as an opt-in research candidate, with a joint GPS/IMU covariance and tests including continuous small aircraft corrections. The saved-data forward-uncertainty-review.json reports angle sigma 0.010193 radians (about 0.58°), while preserving all original failures and recording no science fit. The fresh matched pilot uses a new source freeze and seed; it remains unrun and requires a new compute budget.

What the assumption review found

Five archived route windows were reviewed using the already declared smooth-path interpolation. They contain sparse public position observations; frequent GPS and IMU readings remain simulated assumptions. Reported height and clock references are unconfirmed. These calculations do not measure GPS accuracy or establish what happened between observations.

The existing 20/40/60-minute IMU-turn control passes the two-minute buffered assumed-motion check on Frankfurt–Johannesburg and Seattle–Keflavik. It fails on the selected windows for the other three routes. Those failures remain recorded; the times and motion limits were not adjusted to obtain a pass. Even a passing assumed-path check does not certify the reconstructed orientation or a usable scientific recording.

The physical correction imposes consistency between the wind vector, ground velocity and a slowly varying airspeed. Its fixed constraint width is 2 m/s and its knots are 15 minutes apart. There is no independent airspeed or aircraft-heading measurement. A constrained fit therefore depends on these smoothness assumptions, rather than deriving airspeed uniquely from GPS.

For a separate diagnostic, prescribed wind was subtracted from the assumed ground path to calculate its implied airspeed. Values sampled at 15-minute knots were interpolated without optimization. On Frankfurt–Johannesburg, the absolute mismatch has a median of 2.95 m/s, 95th percentile of 11.41 m/s and maximum of 20.68 m/s; 79.7% of sampled points are within 6 m/s. The other four windows have maximum mismatches of about 23–49 m/s. This calculation spans supported path points, including intervals that later preprocessing might exclude. It is not the optimal airspeed spline, a fitted model residual, or proof that the physical correction fails. It shows why assumed interpolation roughness and airspeed smoothness need a matched test.

All 65 frozen wind/airspeed design states were also checked against their speed constraint. None remains within 6 m/s at every sampled point on any of these assumed paths. The finite grid intentionally includes independent wind/TAS corners and off-constraint combinations; that result is not an identifiability score. No state was removed, tolerance widened, nuisance model changed or science cutoff relaxed. A future review should decide whether a geometry-only envelope over jointly credible states better represents the experiment, using evidence beyond which states produce attractive model separation.

The concrete proposal

The frozen proposal uses the supplied public LH572 FRA–JNB window in simulated-frequent-GPS mode. It keeps the current 20/40/60-minute control, fresh development seed 600900, three model truths and two scenarios: wind with mixed IMU bias, and wind with bias, correlated noise and temperature effects. The three candidates make 18 evaluations of six generating conditions. They share RNG streams for matched comparisons and are not 18 independent null draws.

The initial completed stage is three evaluations of the same rotating-globe/wind-plus-bias condition, one per candidate. Its purpose is to measure current runtime, diagnostic storage, exclusions, parameter-boundary behavior, forward-axis uncertainty and available model-pair information. The remaining 15 evaluations need a subsequent budget decision using those measurements. Zero bootstrap keeps this a development diagnostic; it cannot establish a rare error rate, justify a winner or replace calibration and independent validation.

Use at most two worker processes, each with one numerical thread. Historical 7–12-second flight times are not a quote for the larger nuisance envelope and two-path fits. Successful-fit runtime and storage remain unmeasured. No schedule search is included. True mount-slip controls, other routes, the original sparse-fix lane, larger wind/airspeed departures and complete nonlinear bootstrap remain separately scoped follow-ups. In particular, the mount candidate cannot be assessed for slip detection merely from wind/no-creep cases.

Frozen technical record and reproduction

plan.json is a sharded-flight-campaign-1 plan with source/configuration/policy/one-thread environment and exact candidate jobs frozen. assumption-review.json records every reviewed route, input hashes, turn-mask failures, all 65 speed-state diagnostics, task IDs and the provisional assumptions. Their original preparation declarations remain immutable; subsequent approval covered only the first three evaluations, now spent. run/ preserves their journals, raw sessions, preprocessing arrays and partial export. processing-review.json records the saved-data forward-reference audit from analysis/tests/audit_airline_pilot.py; it generated no flights, science fits, SVD calculations or search. The partial export is not holdout evidence. Preparation helper: analysis/tests/prepare_airline_pilot.py; six controlled checks are in analysis/tests/test_airline_pilot_preparation.py and passed in 1.07 seconds. They fail if preparation invokes a flight fit or SVD, exercise smooth/jittered paths, verify the matched prefix and reject an unsafe selected schedule. The preparation itself took less than a second.

The completed pilot's scientific source was d5d63a6eec604ac2ba02cb39648767a914841f3c05bb8790eeb1ef442f95cb42. The required worker environment hash is 20ef393862e7723ba90e825b625826081cde501e4d9c944f2dbff29526ebdd69. Plan hash: 74b7de930a28daeb5dc56884f1baefc990a32c1bcc68b285ea33ef6c4a75bfe7. The preparation script and archived inputs have their own hashes inside the frozen configuration. Earlier core/wind proposals remain historical and were not overwritten.

That source hash identifies the completed pilot. The explicit smoother trajectory extension subsequently changed the current implementation hash to 3b42547d5a2ffa2a2ef5d975382b4b1b2b0a762fe831493d40cb1b7f469cb682; the old plan is now historical and cannot run under this source. Preparation/diagnostic helpers: audit_roll_proxy.py, forward_smoothing_diagnostic.py and prepare_smooth_route.py in analysis/tests/. Controlled checks for the new helpers and trajectory compatibility, the existing core pipeline, processing audit and original preparation passed 51 tests in 2.39 seconds. This is focused software evidence, not a full-suite or campaign result.

roll-proxy-review.json records the scalar cause audit. matched-smoothing-review.json preserves the first diagnostic; matched-smoothing-support-review.json is the subsequent gap-support-hardened result, using 4,187 windows. smooth-trajectory.json declares the explicit alternative; smooth-route-review.json records its geometry-only comparison. Original failed flight records, plan and preprocessing audit remain intact.

The initial authorized run used this command with --detach; its budget is now exhausted. This is a reproduction record, not authorization to generate new attempts:

/home/sweisman/venv/bin/python analysis/tests/sharded_campaign.py run docs/airline-pilot-preparation-20261006/plan.json --output docs/airline-pilot-preparation-20261006/run --workers 2 --attempt-limit 3

An explicit detached launch can survive terminal disconnection, following the runner instructions. The execution environment must allow child processes to persist after command completion. The output run/ now exists and its status is budget_complete, with three preserved failures. Do not use --attempt-limit 18 without approval of the additional work. No runtime ceiling is encoded in this proposal; a strict wall-time budget would need an execution deadline before launch, not an invented runtime estimate.