Next gate: noisy calibration and later-position prediction

The fixed-error adapter and efficient work accounting are software-qualified. The next gate tests the complete nonlinear estimator. It must not substitute component checks for accurate withheld predictions or calibrated uncertainty.

Match the generated input to its error model

Keep recorded-increment errors and rate-point errors separate. The earlier failed noisy study generated continuous white noise and integrated changing bias directly into increments. Preserve that study and its results.

For rate samples converted by the trapezoidal rule, generate noise at the sample points first. Neighboring intervals share an endpoint. Sampled changing bias must also be integrated by that same rule; it is different from an exact continuous-time bias integral. Qualify this independent generator against closed-form sample covariance and prefix-reproducibility controls before fitting. Point variance is an explicitly supplied squared-rate quantity, not a hardware specification or an automatic conversion from continuous white-noise PSD.

Reserve all generated random variates before drawing. Use separate named streams for each sensor, white noise and changing bias. Extending the later record must preserve the earlier generated inputs exactly.

Qualify the complete prediction path

First qualify short generated cases with all 26/27 calibration columns, each physical candidate, both processing assumptions and both input laws. Fit only earlier noisy measurements, freeze covariance from their declared midpoint anchor, and preserve Gaussian normalization. Keep raw forecasts separate from earlier-error-conditioned forecasts. Score later physical truth only after both forecasts and their calibration binding have been saved.

Distinguish prediction targets explicitly. A future receiver measurement includes receiver white noise, persistent offsets and drift; a noiseless physical position does not. Earlier receiver errors contribute to both conditioning problems, but must not be added to the physical-position target. Carry the receiver state continuously while removing only its future observation loadings for that target. Hold future receiver transformations at the last earlier value under a named policy; do not derive them from withheld coordinates. Keep the training covariance prefix exactly unchanged. A singular future innovation is unsupported by the current full innovations interface and must fail without a variance floor.

Report numerical completion, actual prediction errors, unidentified directions, active bounds and unsupported uncertainty separately. A linear correction outside the permitted calibration box receives no unrestricted covariance. Intervals from an interior linear approximation are diagnostic until nonlinear coverage is independently tested.

Then freeze a finite long-duration study. Keep the existing 10 m horizontal / 30 m height RMS accuracy screen, along with its maximum-error checks. Reserve the exact starts, seeds, source/environment versions and native/packet work before launch. Do not increase noise allowances or calibration bounds after seeing outcomes. No aircraft measurements or protected flight groups enter these software gates.

The initial source-law gate permits at most 24 new deterministic checks plus affected noise regressions, one thread, zero native builds/calls and no optimizer starts. Random samples occur only in explicit finite test fixtures with a precharged variate count. The native fitting matrix and long-run resource definition require their own subsequent freezes; this page does not release an unbounded campaign.

Training adapter · Work accounting · Earlier noisy failures · Prediction protocol.