The tighter acceleration bound survives native checks
All six longer generated controls passed. Native predictions, all 27 derivative columns and all shared coefficient arrays remain unchanged. The known physical probe stays inside each factored remainder bound, and the original acceleration Taylor remainder is retained exactly.
The remaining unshared position bounds shrink by roughly 3–19%, depending on axis and candidate. The 92–93% reduction in the acceleration-change radius does not translate directly into a similar reduction in complete position uncertainty: other terms remain. Complete prediction widths have not yet been recomputed.
The study used its twelve native calls, six enclosure calls and 36 independently verified exact supports, finishing in 94.1 seconds. There were no LP proposals, original aircraft reads or observed fits. The conditioned bound applies only to its hashed training region, not the entire physical bounding box.
Next, test all original eight sectors in each of six model/processing groups under a separate finite freeze. Keep the unchanged all-axis, all-group prediction target. These controls do not establish real aircraft error coverage or clear the gate for observed fitting.