Native prediction adapter: what transfers and what does not
The existing native adapter preserves all 27 nuisance columns and charges native calls before execution. It accepts a supplied training-fitted center, uses the native value and derivative kernels, and does not read Group-1 attitude. Its earlier six short-path controls checked interfaces, derivatives and training-prefix agreement. That completed study remains frozen.
Two requirements prevent simply attaching the newly qualified uncertainty envelopes to those saved outputs:
- Approximation error: the old adapter reports the largest discrepancy among supplied axis probes. That is sampled local information, not a bound over the full nuisance region. The new envelope requires a declared global training error norm and future coordinate error bounds. Missing bounds produce an unsupported result; the sampled maximum cannot be substituted.
- Coordinates and covariance: the old diagnostic builds dense covariance factors and converts coordinate errors to metres using the receiver latitude and height at each epoch. This is acceptable for its descriptive residual report. A new withheld-prediction envelope must instead freeze its coordinate representation, error family and transformations before evaluating withheld positions. Future observations cannot determine the claimed envelope.
The old prediction worker initializes and fits from training observations and checks that their predicted prefix agrees with continuous propagation. These are useful existing safeguards. They do not qualify uncertainty from fitting, covariance estimation or nonlinear approximation. A new adapter must preserve those safeguards while exposing training-only dependencies to the structured forecast and explicitly bounded nuisance calculation.
No old fit is relabeled, no frozen worker is changed and no native evaluation was needed for this source audit. Native uncertainty qualification remains open. ESA receiver-only selection and rate-format controls are independent prerequisites that can proceed without weakening this gate.
Source references: lll.ilvis0_uncertainty_adapter,
lll.ilvis0_uncertainty_geometry, ilvis0_prediction_covariance_worker.py.