The large linear examples fail the full training check
All twelve fixed probes completed correctly. None satisfied the nonlinear training constraints. Their largest earlier-position changes were about 114–1,809 m, compared with the unchanged 1 m allowance.
This is useful: the large spreads in the simplified linear diagnostic have not been demonstrated as genuine nonlinear prediction ambiguity. The proposed calibrations can cancel in the linear formula while failing in the full model. The later displacement played no role in rejecting them.
These are generated-data checks around the saved reference calibration. They do not measure aircraft error, prove that every nonlinear prediction is narrow, or permit unexamined calibration regions to be discarded.
The next saved-data audit will include the 24 shared source terms that the linear-only diagnostic deliberately omitted. It will compare those predictions with the already saved native values. Only matching coefficient maps may be reused; old training-conditioned error bounds do not automatically apply to these new points.