From software coverage to aircraft predictions
The frozen generated-data qualification is complete. All eight checks pass; the six 512-realization coverage controls each have an exact one-sided 99% lower bound above 0.95. No archive measurements were used. These results qualify the uncertainty calculation for its stated Gaussian family and bounded mean errors, not those assumptions for actual aircraft observations.
Native adapter qualification
- Audit the existing native uncertainty adapter's dependencies. Training initialization, calibration, covariance assessment and Jacobians must never depend on withheld positions or fused navigation. Keep every calibration column, including unidentified and boundary directions.
- Express native predicted positions, training design and future contrasts in the same declared receiver coordinates and covariance convention. Preserve covariance time origin/span and integration state at the split. Check these dependencies against independent known-answer controls before archive use.
- Establish an explicit nonlinear approximation allowance over the complete parameter region used for prediction. Local axis probes or sampled extrema are diagnostics, not a global guarantee. An unsupported bound must produce an abstention rather than a narrow interval. Test deliberately inadequate mean models and near-boundary cases, alongside exact controls.
- Freeze one shared calibration/error reference and a finite covariance family for all physical candidates. A candidate cannot estimate a more forgiving scale from its own poor residuals. Distinguish externally supplied assumptions from error properties assessed on training data.
- Freeze representative retrospective NASA splits, adequacy criteria, input dependencies and a separate finite numerical allowance before evaluating withheld observations. Report horizontal and vertical prediction errors; no good horizontal score excuses an implausible height path.
The source, selection and original findings remain unchanged. The earlier incomplete prediction trial remains historical development evidence. The completed software study cannot restart or supply another study's allowance.
Independent prerequisites
If a global native approximation bound or an observational error assumption is unsupported, dependent prediction fits stay blocked. Software investigation can continue, as can the roadmap's ESA eligibility and rate-adapter prerequisites. Those use source integrity, timestamps and receiver trajectory, with no model outcome inspection. Recording 187 remains development material; all other ESA groups retain their confirmation reservation. No archive-wide fitting follows automatically from this software result.