Calibration cross-effects explain most of the sampled mismatch
Adding the 24 shared calibration terms improves agreement for all twelve fixed probes. These terms retain effects such as a calibration offset acting through an uncertain orientation; the simpler linear diagnostic omitted them.
| Processing case | Largest linear-only mismatch | After shared terms |
|---|---|---|
| Zero removal | 845 m | 48 m |
| Profiled removal | 3,955 m | 216 m |
This explains much of why the simplified linear formula admitted calibrations that the full model rejected. The remaining mismatch is still material. The next audit will use the already saved derivatives at each probe to examine that remaining curvature and identify which parameter combinations matter.
These are comparisons at generated points, not aircraft errors or confidence intervals. No earlier training-conditioned error bound was applied to a point outside its stated region. Nothing here proves coverage of unexamined calibrations or changes the original 53-stretch findings.