Long-history prefix comparison
The short comparison passes twelve supported cases and preserves two source-cap rejections. The next comparison must use the six existing, unchanged 96,000-packet generated histories and their saved training-conditioned source references. This is numerical-method development, with no aircraft fits or new flight selection.
Inputs and comparisons
Use both processing treatments for all three candidates. Preserve the original configuration, calibration box, expansion, source reference and five-epoch training proof. Query the common domain at 290 and 438 seconds. A changed proof must not inherit another proof's cached state.
The completed remainder study already preserves an original cold report at 438 seconds for every case. Reusing it can avoid six expensive repeat calculations, but requires a compatibility certificate before launch: identical inputs, time, reference, proof, domain, original cold code, numerical dependencies, compiler and runtime environment. The changed orchestration wrapper must be accounted for explicitly; a matching filename is insufficient. If compatibility cannot be established, do not reuse that reference or silently substitute a different one.
Calculate a new original-cold reference at 290 seconds. Compare both complete batched reports with their corresponding cold reports, retaining all remainders and shared/physical coefficient columns. Independent point/Jacobian queries can check the centers and 51-column designs. Keep the existing 1e-5 m software-oracle allowance; it is not a sensor specification or a prediction-accuracy target.
Cost evidence and finite preparation
The six saved 438-second cold bounds took 6,601–6,677 seconds each under their recorded environment. They must not be repeated merely to rebuild documentation. Use those measured costs to freeze a separate allowance before launch. A new 290-second cold bound plus a two-query batched calculation is expected to take roughly three hours per case; this is an estimate, not a reserved budget.
The preparation should record six identities, at most four workers with one numerical thread each, two generated point calls per identity, one new cold call, one batched call and two reused-reference comparisons. With 96,000 packets, reserve at most 192,000 actual batched advances per identity, including both canonical and tail work. Record source-copy, query-floor, force-profile and initialization work separately. Retain the actual-work counts and complete failures.
Freeze per-call and whole-study time limits, peak-memory limits and the final worker allocation in the executable definition. Do not launch from this prose estimate or spend an old allowance. No retries or expansion follow a timeout.
Measure new wall time and peak memory. Historical timing is not a controlled same-run speed benchmark; report its limits alongside exact packet-work counts. Pass/fail depends on complete numerical equivalence and bounded resources, not on obtaining an attractive speedup.
Passing would qualify long-history reuse for these controls. It would not finish the complete-cadence training/error connection, whole-period uncertainty coverage or observational error-law assessment. The 53-stretch results and protected ESA confirmation groups stay unchanged.