Saved prediction with joint rotation relationships

The 24 joint-product controls pass on all 1,536 sealed physical probes. Local prediction tests preserve source blocks, scale, origin, remainders, all original coefficients and exact endpoints. Their instrumented support protocol reserves six proposals / 24 exact support evaluations per nonempty case.

Use all 48 completed training-conditioned sectors: eight per candidate/processing group. Preserve original physical bounds, training observations and constraints, native prediction/Jacobian, source reference and every unshared remainder. Reuse the sealed generated inputs and native results; no integration or new physical oracle is needed.

Construct the qualified 129 joint-product inequalities with the unchanged dimensionless 1e-8 numerical-control allowance. Keep all 51 original coordinates; the 27 additional coordinates represent arithmetic products, not new physical calibration parameters. The component's parent-reference binding and the full training-reference binding must remain distinct and verified, without rewriting hashed component metadata.

Pad the original inequalities with 27 zero coefficients, then append the new constraints. Start from the previously contracted domain and qualified product intervals. Allow at most eight outward interval-contraction passes per identity. Preserve empty regions, failures and unsupported results. A provably empty necessary region may exclude a sector only if it does not contain known truth.

Compute future supports with every original coefficient unchanged and zeros for new arithmetic coordinates. Each signed LP proposal reserves three exact support evaluations, including fallback and previous-certificate comparison. Reserve six further exact supports for independent endpoint verification. Save prediction before checking known future truth; do not tune from the outcome.

Freeze 48 identities / 6,192 plane constructions / 384 contraction passes / 288 LP proposals / 1,152 exact support evaluations maximum. Empty cases consume no LP allowance; unused work is not transferable. Each LP has a one-second cooperative allowance. Four workers use one numerical thread each, with 480 cooperative seconds for the complete study. There are zero native predictions, integration bounds, new physical-oracle calls, original-aircraft reads, observed fits or retries. Charge before execution and preserve interruptions.

All eight sectors must pass per group. Report their complete union, comparison with the immediate predecessor and original projected baseline. The engineering target remains complete widths at most half the projected baseline on every axis of every group. A failed group blocks full qualification. This is a generated-control comparison, not calibrated aircraft uncertainty or a new Earth-model finding. No observed fit follows from component checks alone.

Publication copies sealed evidence without recomputation and reconciles every hashed reservation event against its task and global allowances.