Qualify propagation conditioned on training acceleration

The saved acceleration projection passed its all-group engineering target. Next develop a separate propagation module; preserve all completed sources.

Certificate consumer

Create a typed, immutable consumer that independently verifies all six exact dual certificates, outward endpoints, complete 51-coordinate input identity, reference/source bindings and contracted domain. Recompute the change radii from verified endpoints. Reject changed coefficients, shifts, negative duals, source/reference/region mismatch, or a physical domain outside the certificate. No LP, truth value or future observation enters this consumer.

The certificate describes the intersection of the saved training inequalities and box, not every calibration in the box. Every bound produced with it must carry that restriction and its digest. Never reuse such a bound in an unrelated training region or as a global bound. All 51 coordinates and existing remainder caps remain. Certificate verification itself must have a finite, precharged support allowance in subsequent studies.

Separate native candidate

Keep the original source-bound reference and native center/tangent definitions. Replace only the full-box acceleration change radius and component ranges with the smaller independently verified feasible-region bounds, when applicable. Retain the original acceleration remainder: a tight change radius does not justify shrinking that separate Taylor remainder. Other physical domains, clock allowances, force transforms and shared coefficient blocks are unchanged.

Qualify known-answer controls with both a point region and genuinely correlated regions. Check rejection of wrong bindings/domains, containment at physically feasible probes, unchanged native centers/shared coefficients, retained remainder allowances and exact accounting. Numerical tests do not establish full-program roundoff or physical observational coverage.

If the consumer and propagation pass, freeze a finite generated-case comparison across all original eight orientation sectors and six model/processing groups. Construct conditioned bounds from training constraints only; save predictions before checking known future truth. Preserve the original all-axis, all-group twofold width target and all failed sectors. No observed prediction or fit follows unless the full required method gates pass. Record native/bound/LP/support work before launch; this plan itself launches none.