Saved prediction with shared vector-norm bounds
This comparison is blocked until the separately frozen 24-case norm qualification passes. Apply only its source-verified vector constraints to every case of the completed calibration-consistency comparison. Preserve all 51 coordinates, physical bounds, error allowances and original sectors. Reuse sealed generated inputs, training evidence, native tangents and nonlinear bounds; perform no new native integration or aircraft fitting.
Add 182 necessary inequalities for seven verified error vectors. Their 26 integer directions provide outward supports of a Euclidean unit ball; zero-cap blocks use the existing zero-error normalization convention. Do not assign this structure to the componentwise velocity/clock block. Keep all previous necessary constraints and valid certificates. Recontract domains using at most eight outward row passes, then reserve six support proposals per included case.
Freeze 48 identities, sources, environment, input bundles, qualification evidence and baselines. Maximum: 288 LP proposals and 1,152 exact support evaluations, including fallback checks and the independent endpoint audit. Zero native calls, packet visits, native bounds, observed fits or original aircraft reads. Four workers use one thread each, with a 300-second cooperative allowance. No retries, extensions, process signals or protected ESA outcomes.
Save revised constraints and future predictions before inspecting known physical truth or future outcomes. Every in-sector truth must remain feasible and contained. New ranges must not enlarge; retain previous valid certificates as fallbacks. Publish all 48 outcomes and complete unions across sectors.
The engineering target remains unchanged: at most half the projected common-acceleration baseline width on every axis in every model/processing group. Report the immediate predecessor comparison as well. Correctness alone does not pass the usefulness gate; preserve an unmet target. The single local saved-case smoke test precedes the freeze and permits six support proposals, with no native calls. Whole-program numerical treatment, recording-scale usefulness and observational-error adequacy remain open requirements.