Can training narrow the calibration box?

Use the six completed generated-data cases from the full-history prediction test. Verify their saved result and training-region hashes. Project each of the 27 physical parameters while retaining all 51 physical and shared coordinates, every training inequality and the original calibration limits.

Each signed coordinate bound needs a numerical proposal and an independently checked exact certificate. A failed proposal retains the original box. Narrower coordinate intervals do not replace the coupled constraints or establish that every point in their Cartesian product is feasible.

Limits: six identities, one numerical thread, 600 cooperative seconds, no retries; 54 proposals and 162 exact checks per identity, totaling 324 and 972. Each proposal has the shared one-second solver limit. Charge before execution in indexed journals. There are zero native calls, new trajectory bounds, aircraft reads or model fits. Preserve the earlier unsuccessful prediction ranges and abstention.

The result determines whether a safely narrowed parameter box is worth testing in a separately frozen position calculation. It cannot qualify prediction accuracy or an aircraft error model.