Training and forecast qualification

This software test checks the complete earlier/later prediction path without reading aircraft recordings. It follows the revised prediction protocol.

Fixed controls

Generate constant-speed equatorial tracks from closed-form kinematics, separate from the native integration and derivative code. Include the specified-field globe, zero-field globe and stationary disc. For each, test unmodified angle increments and a known 35% removal of the conventional background vector. Use both nominal 200 Hz periods and alternating 4.9/5.1 ms header periods. Every history lasts 24 seconds, with 4,800 packets and 23 receiver epochs at seconds 1 through 23. The first 15 epochs train; the last eight are withheld.

Initialization sees training positions and their supported IMU prefix only. Use the existing candidate-specific coordinate initialization, with force magnitude computed as summed velocity increments divided by actual elapsed time. Fit all 26 nuisance parameters, plus the removal fraction when profiled, under the existing ±1°/hour offset and other calibration bounds.

Use the existing bounded trust solver, its fresh stationarity/position checks and two fixed starts: all zero, then alternating ±0.001 normalized coordinates. Select a primary by smallest converged training cost, with earliest start as the exact tie rule. Withheld positions never select starts or calibration. Continue the native trajectory without a reset at the split.

Limits and checks

There are 24 fit identities: three candidates × two removal cases × two clocks × two starts. Each permits at most 600 training tangent calls and one complete forecast, for 14,424 calls in total. Every native call and its conservative packet/query work are reserved before execution. No retries or original reads. The study uses one worker and four numerical threads, with a one-hour cooperative wall ceiling. Source, inputs, compiler and numerical environment are frozen before study execution.

Require a converged training solution, correct accounting, all active columns, and training and withheld position errors below 1 mm RMS per axis and 1 cm maximum per axis on these noiseless controls. Compare both starts' objective values to the zero-cost analytic target, allowing at most 0.01 in the summed squared training residual normalized by a declared 1 cm position scale. This scale defines a software check, not a receiver accuracy claim. Weak calibration coordinates need not match a unique parameter vector. These short controls do not establish long-record conditioning, instrument uncertainty or observational adequacy.

Failures stay published. They block dependent archival fits until a separately qualified procedure is available; they do not alter earlier studies or consume reserved ESA confirmation data. Unit tests can check the implementation before the separately frozen qualification run.