Prediction tests with known measurement errors

Purpose

The completed long software controls used noiseless measurements. The next test adds known receiver errors, gyroscope noise, accelerometer noise and slowly changing sensor bias. It asks how the existing training-only estimator behaves when these errors occur together.

This is a generated-data robustness study, not an estimate of the archive's actual noise or a calibration of scientific confidence. The existing receiver weights remain fixed: they must not be enlarged because a candidate fits badly. A receiver-only weight does not become a complete IMU prediction-error model.

Input qualification

Use the shared lll.measurement_noise generator and lll.airquantumgrav_noisy_inputs constructor. Apply errors before initialization and fitting, including the first receiver height used to anchor gravity. The fitter receives noisy increments and earlier noisy positions only, without true calibration, attitude or initial state.

White sensor errors integrate with the square root of the actual sample interval. Slowly changing bias uses a stationary Ornstein–Uhlenbeck process; its interval integral and endpoint are generated jointly. Do not approximate an entire interval by its endpoint bias. Retain their correlation and support uninterrupted continuation.

Receiver errors preserve the supplied cross-axis covariance, temporal correlation and fixed drift origin. Use independent named random streams for receiver, gyro, accelerometer, white noise and changing bias. Extending later samples must leave the earlier generated noise unchanged. Reject invalid covariance or timing without clipping or repair.

The input checks include high-precision integral identities, independently written covariance expectations, irregular timing, zero noise, exact prefix reproducibility and all six model/processing combinations. These checks qualify input generation; they do not qualify noisy fitting or aircraft uncertainty.

Declared stress settings

These are development assumptions, not measured instrument specifications. They are fixed before the noisy fitting results are examined.

Keep the existing constant calibration bounds as well. Changing bias and constant offset are distinct. Match the same sensor-noise realization across starts and receiver-weight cases; preserve both receiver policies.

Fitting and scoring

Reuse the qualified wider-precision training/forecast path. The matrix has three generating physical candidates, two processing cases, two receiver policies and two starts: 24 identities. Each candidate is tested against data generated under that candidate; this is not an observed model-selection campaign. Keep the twenty-minute turning/climbing trajectory, 800 training positions and 400 later query epochs, with all 26/27 nuisance columns.

Generate later truth positions only after the continuous forecast is saved. Score against known physical truth, without adding later receiver noise to hide reconstruction error. Use the existing observed-prediction accuracy screen: horizontal RMS 10 m, height RMS 30 m, maximum horizontal error 50 m and maximum absolute height error 150 m. Report numerical completion and physical accuracy separately. The earlier noiseless software targets and results remain unchanged.

Before execution, qualify the actual fitting adapter and reporter, including absence of later-position influence, noisy initialization, complete derivative columns, native accounting and preservation of failures. Input-generator tests alone do not release this study.

Separate finite allowance

Freeze the final definition before execution. Allow 24 starts, at most 200 fitting evaluations and one forecast each, four single-thread workers and a cooperative eight-hour ceiling. This is a new allowance, not unused work from a completed study. Native-call ceiling: 4,824. Packet/query work ceiling: 931,411,200,000. No retries or transfers between starts.

Per identity, charge at most 360,000 constructed intervals, 360,801 generated samples before scoring, 1,200 later truth samples and 6,484,812 normal variates. Charge before work and preserve failed or interrupted starts. Archive original integer measurements, observed fits and protected data are not inputs.

Results can expose a failure of the working error assumptions or estimator. Do not change bounds, noise amplitudes, stopping checks or accuracy limits after seeing an outcome to make the study pass. A failed gate blocks dependent confirmation work; separate diagnosis remains permitted.

The 53-stretch baseline and seven protected ESA confirmation groups stay unchanged. These tests do not identify a rotation field's cause or establish a solar-versus-sidereal distinction.

Error diagnostic · Prediction protocol · Roadmap.