Training-only prediction interface controls

Seventeen component tests pass for the new interface between supplied native predictions/derivatives and the qualified structured uncertainty calculation. It performs no integration, fitting or archive reads.

The checks cover a nonzero fitted center, complete four- and 27-parameter boxes, unidentified directions, affine coordinate transformations, a two-member covariance family and an independently bounded quadratic mean error. They also check that a bad candidate cannot enlarge the shared scale, a missing covariance alternative cannot disappear, and a bound cannot be reused for another model, center, box, parameter order or numerical error configuration.

Future observations are absent from the API. Coordinate transforms are frozen configuration; the interface neither recalculates them from withheld positions nor silently wraps longitude. Native Jacobians use the full training-noise dimension. Missing global approximation bounds produce unsupported envelopes; sampled local probes are not accepted as a global-bound declaration.

This qualifies interface calculations on supplied mathematical inputs. It does not establish a global native remainder bound, qualify the source of the native predictions, or calibrate the receiver-error family on aircraft observations. Every output remains a scientific abstention, including supported mathematical envelopes. These gates remain ahead of a new observed prediction study.

Implementation: lll.ilvis0_prediction_envelope_adapter. Checks: analysis/tests/test_ilvis0_prediction_envelope_adapter.py.

Interface design · Mathematics · Native source audit · Next gates.