Complete prediction comparison with a common force transform

The common matrix, clock/integral, native and typed 51-column interface gates pass. Compare the full new construction with the two preserved 7.2-second baselines: the ordinary common-acceleration comparison and its stronger coordinate-projected follow-up. Neither baseline is changed or relaunched.

Reuse the exact six generated 1,600-packet inputs and all eight original orientation sectors. Preserve all 27 physical bounds, both processing cases, the six training times (0.8, 1.4, 2.4, 3.6, 4.8, 6 seconds), future epoch 7.2 seconds, 1 mm generated observation allowance and 10 micrometre arithmetic margin. These allowances are engineering inputs, not measured receiver accuracy.

One fixed 24-coordinate common-error vector spans all epochs; the 51-column training system retains every physical direction. Future receiver truth and the known-control common errors are evaluated only after the training evidence and future envelope are emitted. Preserve pure training exclusions, all sectors, known-truth retention and the existing 1 micrometre native-centre limit.

Correctness requires all 48 identities and all six complete sector unions. The preregistered engineering usefulness target is at least a twofold width reduction on every axis in every case relative to the projected 7.2-second baseline. Also report the ordinary baseline comparison. This is a method development target, not a detection threshold or sensor accuracy claim. Do not relax it after seeing results. The new method uses row contraction; the stronger baseline additionally used coordinate projection, which remains explicit.

Freeze 48 starts, at most three native calls and eight bounds per identity: 144 native calls and 230,400 native packet visits in total. A bound reservation counts one complete exact-clock pass plus one propagation pass, at most 3,200 steps per bound: 384 bounds and 1,228,800 combined reserved steps. This new two-pass accounting does not reinterpret any completed study's allowance.

Four workers use one numerical thread each. Allow 5,400 cooperative seconds overall and 120 per bound, shared between integral construction and propagation. Expected runtime is roughly 30–60 minutes from the preceding studies. No retries, extensions, signals, original aircraft reads, observed fits or protected ESA outcomes. Verify compiler/environment identity and freeze all inputs, sources, baselines and rules before execution.

Only verified evidence may guide the next recording-scale feasibility step. Native arithmetic treatment, actual observation-error adequacy and independent confirmation remain open; this study creates no scientific model preference.