Check propagation using training-derived acceleration bounds

Prerequisites: the six saved acceleration projections passed exact checks and their all-group engineering target. Local consumer checks reject corrupted certificates and mismatched sources/domains. Twelve local point/correlated-bias controls across all candidate/processing groups preserve native centers, all 51 coefficients and feasible-probe containment; one rejection test also passes. These do not establish observational coverage or complete roundoff bounds.

Freeze six identities, the same unique known-truth sectors as the acceleration projection. Use their sealed generated input, training constraints, reference, certificate and previous future bound. Preserve all completed sources.

For each identity, independently verify six exact support certificates before native computation. Charge those six checks before execution. Reconstruct the original source/reference, require exact identity, and run one separate training-conditioned enclosure. Preserve original shared-error coefficients and Taylor remainder allowances. Only the feasible acceleration-change bounds and physical contracted domain may improve. The result applies to that entire training region, not to unrelated calibrations in its bounding box.

Save the conditioned bound before reading known calibration or future truth. Then call the unchanged native kernel at the original center and one known physical point, retaining all 27 tangent columns. Check the center against the enclosure within the existing 1e-6 m axis limit, unchanged saved center/shared coefficient arrays, native Jacobian agreement, and factored residual containment at the known point within the existing 1e-5 m numerical allowance. Preserve the original acceleration remainder exactly. No optimization or known-value tuning.

Limits: six identities; twelve native evaluations total; six enclosure calls; 36 exact support verifications; zero LP proposals. Each native evaluation reserves the complete input packet count, and each enclosure reserves twice that count. At most 60 cooperative seconds per bound and 600 seconds for the study. Four workers use one numerical thread each. No originals, observed fits, retries or budget expansion. Save failures without replacements.

Qualification requires every identity to pass source/domain, native-center, coefficient, derivative and containment checks. Narrower remainders alone cannot qualify it. If it passes, a separately frozen complete eight-sector/six-group prediction comparison may test the unchanged all-axis twofold width target. No observed prediction or fit is authorized by these component controls.