Check local timing verification in the full simulator
The local mathematical qualification passes all 36 fits, eight direct controls and nine starting-point comparisons. It replaces distant timing-cell midpoints with midpoints inside the existing ±10-microsecond neighborhood. The earlier four-fit trial remains failed.
Frozen question and inputs
Run the same four exposed generated-data cases from their original declared starts, not their saved fitted answers. Use seeds 281474977010692 and 281474977010693, the point-error law, specified-field globe and the original profile setting, with starts zero and one. Preserve the two originally passing controls and two originally failed labels. Require byte-identical original input, training and error-operator bindings.
Only the qualified local timing verification changes. Preserve all 27 nuisance columns, physical bounds, mean/noise model, packet clock, near-zero rule and forecast screens. Keep numerical limits: relative cost 10⁻⁸, stationarity 10⁻⁴, position-change RMS 1 mm and maximum 1 cm. Actual fresh evaluations must verify every retained timing cell; no cell is dropped after seeing its result.
Separate finite allowance
This is a new four-fit allowance, not a retry or extension of the closed trial. Freeze source, environment, selection, inputs and budgets before launching. Use four workers with one numerical thread each, at most 200 residual/Jacobian evaluations per fit including final checks, 620 bounded linear operations per fit, two native builds per worker and a one-hour cooperative ceiling. No retries. The unchanged shared accounting formula caps native calls at 239,816 and packet visits at 1,506,051,012. Interrupted starts remain charged; unused work closes.
Reuse the shared definition, evaluator and reporter; do not copy worker trees. First pass source-bound software tests. Independently audit saved traces, charges, original bindings and both starting-point comparisons after execution. Publish every failure without refitting during packaging.
Decision
All four fits, raw forecast screens, original binding replays and both paired start checks must pass before preparing a separately frozen 48-case optimizer qualification. A failure blocks that dependency and is preserved for diagnosis. Neither result reopens calibration automatically. Protected aircraft groups and the original 53-stretch findings remain untouched.