Comparing two wind corrections

This preparation describes the source committed as db28900. Subsequent magnetic-ambiguity development changed the implementation and policy freeze. The proposal below is historical and cannot be executed under current source without refreezing; no evaluations have run.

A subsequent airline assumption review and smaller pilot freezes the current source and single-thread worker environment, with an initial proposed three-evaluation timing stage. It preserves this original proposal as historical evidence.

The experimental physical wind/airspeed correction is implemented. It describes wind and aircraft movement through the air separately, then checks whether their combination agrees with GPS ground motion. The earlier wind correction instead allows slowly changing sideways pointing coefficients. Neither has established reliable decisions for real IMU recordings.

The next comparison should use the same simulated recordings for both corrections, including wind alone, wind with mixed IMU drift, and wind with drift, correlated noise and temperature. The independent wind-truth generator remains unchanged. Retain all exclusions, fit failures, model-pair information, prior sensitivity, bound hits and runtime; do not select only successful fits. A smaller nuisance space is useful only if it still covers plausible errors.

proposal.json is a preparation-only manifest for a Frankfurt–Johannesburg development window with simulated frequent GPS along the assumed short-interval path. It proposes three paired seeds, three Earth truths, three nuisance scenarios and two corrections: 27 simulated recordings evaluated 54 times. IMU turns at 20, 40 and 60 minutes pass the revised two-minute-buffered maneuver mask on that assumed route; the schedule is not optimized. Its timing checks are geometry checks on an assumed path, not verification of aircraft motion between public fixes or reconstructed IMU orientation. Bootstrap is zero, so this comparison cannot establish rare error rates. Runtime is unmeasured.

Authorized executions: zero. The manifest is not run permission. Obtain a bounded timing and diagnostic-storage budget before execution. There is no final calibration or validation policy. Older proposals remain historical for their frozen sources.

The refreshed optimizer-estimate.json retains ten permitted grid times and an upper bound of 5,571,423 SVD calculations for a three-turn, beam-16 search with both wind corrections. No search or SVD scoring was executed. This bound explains why search needs a separate budget.

Software verification completed with 336 tests passed in 4 minutes 42 seconds. The server's async upload test stalls inside the execution sandbox; the complete suite passed outside it with a ten-minute process deadline and two numerical threads. One dependency deprecation warning was reported. Tests do not establish flight-domain power or error tails.

The physical candidate requires --crab-model wind_tas --design-mode envelope in analysis/tests/research.py; it can be compared with wind in the same invocation. Requested bootstrap also requires --bootstrap-refit nonlinear. Geometry search accepts wind_tas among the worst-case nuisance candidates. Physical priors and bounds are frozen in analysis/lll/wind_tas.py; changing them requires a new policy/source freeze.

GPS alone supplies no independent aircraft heading or airspeed observation. The physical constraint is conditional on the recorded ground vectors, and its gyro bootstrap holds them fixed. Unknown public-track errors remain a separate limitation. The finite design envelope checks stated nuisance examples, not every wind trajectory. See the technical validation record for the numerical assumptions.