Shared-reference orientation-cover usefulness

Repeat the same six generated cases and eight closed orientation sectors used by the earlier two comparisons. Keep their byte-identical 400-packet inputs, original 27 physical bounds, training times 0.2/0.35/0.6/0.9/1.2/1.5 seconds, future time 1.8 seconds, 1 mm training allowance and 10 micrometre arithmetic allowance. No sector may be selected using future truth. Preserve all results.

Prerequisites: the 24-identity shared-native qualification and the 18-identity joint-interface qualification must be complete, passing and hash-verified in the same numerical environment. One fixed source/expansion binding per sector links its six common error coordinates across all training and future epochs. All 33 enclosure columns remain. Keep the fixed physical expansion after contraction; bound the containing hull if the center becomes infeasible.

Use the same outward necessary-constraint contraction and exact certified support method. Save native values/Jacobians, common-error radii and binding, complete training inequalities, contracted joint domain, unshared contributions and future support certificates before evaluating future truth. The actual known-control common-error vector is only a subsequent retention oracle.

Correctness requires every sector containing the known truth to retain its full physical-plus-common-error vector and enclose its future output. Native/bound centers must agree within the unchanged 1 micrometre limit. Publish the union of every nonexcluded sector. The usefulness target remains tenfold narrower than the original broad-box baseline on every axis in all six complete cases. Also report the change from the completed regrouped comparison. Correctness alone is insufficient. Do not relax physical bounds or tolerances to pass.

Freeze 48 sector identities / 144 native calls / 57,600 native packet visits, 384 bound evaluations / 153,600 reserved bound steps, 48 contractions, 288 LP proposals / 576 support certificates and a 900-second cooperative limit. Reserve one worker with four numerical threads. Charge native calls and reserve bounds before execution; no retries, expansions or process signals. No original files, observed fits or protected ESA outcomes are read.

Even a pass qualifies only this short generated usefulness comparison. Longer recording-scale usefulness, whole-program numerical enclosure, actual receiver error adequacy and independent confirmation remain separate gates. All scientific outputs abstain.

Native prerequisite · Joint interface · Previous comparison · Roadmap.