Broader qualification of local timing checks
Conditional plan: this phase cannot start until the four-fit regression qualifies. Its result must be independently audited, published and bound by source and environment hashes. Preparing this plan spends none of the proposed calculation allowance.
The prerequisite now passes: all four fits, both start comparisons and original input replays qualify. The archived publication replays identically, with all 366 receipt artifacts verified. The matrix still needs its own software qualification and freeze.
Question
Does the local timing method also work across both simulated input formats, all three physical candidates and both processing settings? The four-fit gate covers only the point-rate, specified-field globe and profiled-processing case. It cannot answer this broader question by itself.
Fixed matrix
Repeat the previous 48-case qualification using its existing exposed development seeds 281474976810725 and 281474976810726. Keep both input laws (increments and point rates), three candidates, two processing settings and two original declared starting points: 48 fits and 24 matched comparisons. Do not initialize from saved fitted values, change the random draws, select by later errors or introduce new observations. Preserve the four original input, training and error-operator replay checks.
The only scientific-path change is the qualified timing-aware stopping check. Keep the compensated integrator, noise propagation, physical parameter domain, near-zero rule, forecast screens and every nuisance column. Different processing settings retain their respective original parameter lists. Record every timing cell and both successful and failed checks.
Finite allowance and implementation
Use a new definition and source/environment freeze with the shared study runner, explicit solver provider and independent saved-result auditor. Reuse library components; do not copy worker trees. First pass source-bound software tests. The shared binding input carries the four preserved original checks; task identities, computational payloads and limits must match the old 48-case matrix. Validate both 26-column and 27-column processing configurations without dropping their respective nuisance parameters. Shared execution/reporting hooks may be generalized; the seven qualified numerical method bodies remain hash-bound.
The ceiling is 48 fits at 200 evaluations each, including final verification, with no retries. Precharge at most 620 bounded linear operations per fit (29,760 total), with 100 iterations per operation. Keep the previous shared native accounting limits: 2,877,792 calls and 18,072,612,144 packet visits. Verify those totals against the generated task table before preparation. Use four workers at one numerical thread each, two native builds per worker and a two-hour cooperative ceiling. Charge before execution, preserve interrupted starts and close all unused work when the phase ends. No previous allowance is reopened or transferred.
Acceptance and next gate
Require all 48 numerical/physical forecast checks, all 24 paired-start checks and all four original binding replays to pass. Paired checks retain the existing training-cost and reconstructed training-position tolerances, including height; each later forecast separately has to meet the original raw accuracy screens. Report those checks separately rather than treating agreement during training as proof that later prediction uncertainty is calibrated.
Packaging must verify saved evidence without refitting. Publish every result, including failures, then update the roadmap. A pass qualifies this generated-data matrix; it supplies neither an aircraft confidence level nor a calibration threshold. A failed gate blocks a dependent release.
Any later calibration release needs a separate finite plan. Both earlier calibration campaigns remain closed. Protected ESA groups, raw observations and the original 53-stretch findings remain unchanged.