Complete prediction comparison with coupled calibration errors
Reuse all 48 saved orientation sectors from the completed joint-product comparison, under its original six generated candidate/processing cases. No source measurement, native coefficient, reference radius, physical bound, reported failure or previous outcome changes.
Computation
Reconstruct and verify the complete original reference. Evaluate the qualified static map on the previous contracted physical box and its fixed midpoint. Its 24 error terms must remain functions of the same 27 physical parameters. Add the resulting 48 necessary inequalities to all previous constraints; the 27 existing arithmetic-product coordinates remain, with zero coefficients in these new rows. No new physical parameter or observation is introduced.
Before contraction or optimization, independently reconstruct every coupling row and outward upper bound from the saved gradient intervals, center ranges and domain using exact rational arithmetic. This audit performs no support calculations. Retain the declared 3e-10 normalized numerical-value allowance. The earlier 3e-8 derivative comparison tolerance does not enter these rows.
Contract at most eight passes, then compute six signed future supports per nonempty sector. Old certificates may be reevaluated with zero padding. Each proposal reserves three exact support calculations: two in the proposal helper and one old-certificate reevaluation. Reserve six more for the independent endpoint audit. Do not rerun an old baseline audit.
Only after the ranges and audits are saved, check the sealed known generated calibration in its declared sector and the saved future truth. Never use either to choose a center, bound or error law. Keep empty and unsupported sectors and all failed controls without replacements.
Frozen allowance
48 identities; at most 48 source setups, 96 interval-map evaluations, 48 coupling audits (2,304 exported rows), 48 contraction runs with at most eight passes, 288 one-second LP proposals, 1,152 exact support checks and 48 generated-truth audits. Contractions retain all 78 columns and at most 1,024 rows. All work is reserved durably before execution; interrupted work remains charged. Four workers use one numerical thread each, with a 480-second cooperative wall allowance. No retries, native predictions/bounds, original- aircraft reads or observed fits. Completed allowances cannot expand or resume.
Gate and interpretation
All 48 sectors must pass correctness, and every six-group union must retain its known truth. Preserve the unchanged twofold width target on every axis, relative to the original projected 7.2-second baseline. Also report the immediate preceding widths. Partial improvements do not qualify a rollout.
This tests generated prediction ranges under stated numerical assumptions. Useful full-flight uncertainty, real receiver covariance, withheld aircraft prediction and calibrated scientific confidence remain separate requirements.