Complete conditioned prediction comparison

Use all 48 saved column-orientation sectors, preserving their training regions, original physical limits and complete 51 coordinates. Reuse the six generated input cases and their saved native centers, Jacobians and known outcomes. Do not reuse a completed study's allowance or edit its sources.

For each sector, derive acceleration bounds using only its training region: six LP proposals and eighteen exact support checks. Reserve another six exact checks before the typed consumer independently verifies that certificate. Reconstruct the original reference and compute one training-conditioned native enclosure on the same contracted physical domain. Check unchanged centers and all shared coefficients against the sealed previous bound.

Combine the new unshared remainder with the original affine future expression. The smaller of independently valid old/new component remainder bounds may be used, but no allowance may be discarded. Reuse the qualified future-support builder and independent endpoint audit: six more LP proposals and 24 exact support checks, including full-box fallbacks and previous-certificate comparisons. No extra baseline certificate audit is hidden in a task.

The frozen total will therefore be at most 48 identities, 48 native enclosure calls (twice the input packet count reserved per call), 576 LP proposals and 2,304 exact support evaluations. There are zero native prediction calls: saved centers/Jacobians and known controls suffice. Four workers use one numerical thread each, with 120 cooperative seconds per enclosure and 1,800 seconds for the study. No original reads, observed fits, retries or expanded allowance. Implementation and focused accounting tests must pass before freezing/launch.

Save every prediction before reading known truth for containment. Keep all eight sectors in the published union, with failed/unsupported sectors blocking that group. Compare with both the immediate predecessor and original projected baseline. Success still requires at least twofold narrower complete widths on every axis of every group, together with every correctness/containment check. Component improvements alone cannot pass that gate. No observed fits follow without the remaining physical coverage and uncertainty qualifications.