Complete prediction comparison with directional acceleration

The directional static component and all 24 matched native controls pass. This study measures whether the improvement produces useful complete prediction ranges. It uses every original model/processing group and yaw sector: 48 identities, with the same six generated 1,600-packet inputs, six training epochs and held-out prediction at 7.2 seconds. All measurements, calibration bounds, original reference points and 78 enclosure coordinates remain.

For each parent, use its already-certified physical domain and all inherited necessary constraints. Construct one source-bound directional acceleration certificate. Recompute six training bounds and one future bound. Transfer any bound/native-center difference into the remainder with exact outward rounding; center agreement must still be within 1 micrometre. All 51 physical/common coefficients must match the original expression exactly. Preserve the old shared error caps and acceleration remainder.

Add the new training inequalities using the unchanged 1 mm generated observation allowance and saved arithmetic margin. Never remove an old constraint. An independent exact audit must reconstruct all 36 added rows. Apply at most eight contraction passes with the 1,024-row/78-column ceiling. Run six signed LP proposals and independently verify all six exact future endpoints.

Save the fresh range even if it is wider. After all eight sectors are unioned, also report its intersection with the already-qualified 192-box range. Both calculations cover the same physical possibilities, so their intersection is a valid combined range; do not hide either component. Compare both fresh and combined ranges with the unchanged twofold-width target from the original projected 7.2-second comparison. Every axis in every group must meet the target for the combined method to pass. Missing/failed sectors prevent qualification.

Only after bounds or an empty proof are fixed may generated calibration and future truth be inspected. Check every known in-sector control, including its full lifted calibration point and all new inequalities. Retain every outcome. No original aircraft/ESA data, new native predictions, model fits or new noise assumptions are introduced.

Finite allowance: 48 reference setups, 48 directional certificates, 336 propagation bounds, 1,075,200 bound packet-steps, 1,728 independent training-row checks, 48 contractions, 288 LP proposals, 1,152 exact support evaluations and 48 post-range truth audits. Each proposal reserves its two helper support evaluations plus one inherited-certificate reevaluation; each endpoint audit reserves six more. Four workers use one numerical thread each. Each bound has a cooperative 120-second ceiling; the overall ceiling is 2,400 seconds. Expected wall time is roughly 20–35 minutes. No retries, allowance expansion, process signals, observed fits or production promotion.

Passing tests usefulness on these generated histories. Recording-scale bounds, floating-execution enclosure and observational error coverage remain separate gates before aircraft prediction fits can proceed.

Native qualification · Previous complete comparison · Roadmap.