Full-history check of the batched calculation
Run only after all 24 batched native controls pass, including exact agreement with all 348 saved predecessor bounds. This is a separate study, not a retry or extension of the completed time-limited source screen.
Use the same six generated 1,600-packet sources, original calibration domains, saved training constraints and query times. Compute one bound per source under the original cooperative 120-second bound limit. Retain all point predictions, 51 coefficient columns and physical limits. Compare the result with the best compatible saved allowance on each axis. Smaller height allowances alone do not qualify the complete horizontal and vertical prediction target.
Separate finite allowance
- Six identities, one bound each, no retries.
- 38,400 world/source packet-steps maximum.
- Six setups, 36 exact training-support verifications, 9,600 compiled source steps and 9,600 rounded body bounds, reserved before execution.
- Four single-thread workers; at most four total numerical threads.
- 120 seconds per bound and 600 seconds overall, checked cooperatively.
- No full native-trajectory evaluations, optimizer starts, LP proposals, aircraft originals or protected ESA observations.
Freeze code, sources, inputs, compiler and environment before launch. Preserve all failures and charges. Success establishes that the batched calculation completes and keeps compatible reference maps on these source domains; it does not establish useful complete prediction widths, observational coverage or new aircraft findings. Proper source-specific Taylor information remains a separate numerical-development task if the horizontal allowances stay wide.