Saved longer-prediction audit

Verify and diagnose the 48 complete generated sectors at 7.2 seconds using only their sealed results, training inequalities and future evidence. Preserve all six model/processing cases and all 42 columns. No measurements, integrations, parameter fitting or linear-program proposals are introduced.

Independently recompute each of the 288 support certificates using exact rational arithmetic. Reconcile the saved future endpoints with their affine support, unshared remainder and arithmetic allowance. Report position-propagation and lever contributions, and which physical/auxiliary box coordinates narrowed. An unchanged coordinate box does not imply that the joint feasible set failed to narrow: its retained correlations can still improve prediction.

Freeze the saved evidence, implementation, 48 identities and environment. Limits are 48 starts, six exact support checks per identity, 288 checks total, zero new native calls, zero native bounds and zero observed fits. Use four workers with one thread each and 120 cooperative seconds. No retries or allowance extensions. Retain any inconsistency as a failed audit.

This audit verifies internal mathematics and diagnoses this conservative construction. It cannot establish receiver accuracy, calibrated coverage, physical identifiability or a scientific model preference.