Bound combined acceleration from the training constraints

No individual shared-error block dominates the saved widths. Before another matrix refinement, test a combination that the propagation code currently bounds using separate coordinate ranges: the reference acceleration.

The source-bound acceleration has the existing representation A_i(x) = a_i + J_i(x_physical-z) + cap_i gamma_(39+i). The training region already relates all 51 coordinates through receiver constraints. A linear combination can therefore be tightly bounded even when its individual coordinates remain wide. Preserve every coordinate and compute exact dual support bounds for this complete affine expression.

Local checks

Use the qualified affine-support proposer and an independent exact support verifier. Check a known correlated region whose individual coordinates remain wide but whose sum is narrow, signed objectives, zero caps, correct column 39–41 mapping, source-binding mismatch and absent future observations. No known calibration or future outcome may construct the bounds. Known generated anchors are used only after saving them, to test containment.

Reserve three exact support evaluations before each signed proposal: two in the proposer, including its full-box fallback, and one independent verification. Do not repeat the old future-certificate audit inside each task. Check the sealed prerequisite receipts during preparation instead. Test the counters.

Finite saved-only assessment

After local tests pass, freeze six identities: one previously declared known-truth sector per model/processing group from the latest complete comparison. Each bounds three acceleration components in both directions: 36 LP proposals, 108 exact support evaluations total. Four workers use one thread each with a 180-second cooperative allowance. No native integration or bounds, original aircraft reads, observed fits, protected outcomes, retries or budget expansion.

Use the saved acceleration center, Jacobian, caps, source binding and training region. Preserve original calibration bounds and the complete region; this is projection of a combination, not removal of an unidentified direction. Save coefficients, nonnegative dual certificates, exact supports and outward acceleration endpoints before reading the known physical anchor.

For each component derive the largest distance of its interval from the unchanged expansion center. Their Euclidean norm bounds the total acceleration change. Compare it with the source's existing full-box change radius. The engineering usefulness criterion is at least a twofold reduction in that radius in every group. This does not replace the unchanged prediction-width target.

If useful, prepare a separately qualified propagation candidate that consumes the verified observable bounds only where valid for the supplied training region. It must preserve the original reference, all remainder allowances not newly justified, coefficient columns and unchanged native center/tangent outputs. Conditioning must be source-bound; a certificate for another region cannot be accepted. No current native source or frozen evidence may be edited.

This study alone would not produce new prediction intervals, establish physical coverage or clear the gate for observed fits. It tests whether joint training information is lost when later propagation uses coordinate boxes.