Native qualification of timing-aware stopping
Question
Does the analytically qualified method also work through the complete generated-input fitting and prediction path, with the original physical integrator, error model and calibration bounds?
First gate: four regression cases
Use only the already exposed seeds 281474977010692 and 281474977010693, point-observation input law, specified-field globe, profiled processing parameter and original starts 0 and 1. These are the two passing controls and two numerical failures preserved in the closed calibration record.
Regenerate their original inputs and run one new fit from each original declared start. Verify exact input-array, training-observation and error-operator bindings against the saved records. Do not start from the old fitted answers, select a preferred start using forecast errors or change either noise draw.
Use explicit provider hooks through the shared structured-input runner, training protocol and counted fitting adapter. Select the new timing-aware solver only for this frozen study. Preserve the default providers and archived source; do not copy a worker or patch process-global functions.
Keep all 27 active parameters, compensated mean calculation, noise propagation, physical bounds and existing numerical/forecast limits. Record every final timing cell, actual cost/position check and original label. Predictions from both starts must meet the predeclared agreement checks, including physical height. Numerical completion alone is insufficient.
Finite allowance
Four fits, at most 200 fit evaluations each, no retries. The 36-evaluation timing-verification reserve and two near-zero repeats are included in that 200, not added to it. Retain the existing mean forecast and error-propagation diagnostic, including their unchanged operation accounting. Derive construction/native call/packet limits from the unchanged shared recipe before preparation, and freeze their exact totals.
Use four workers with one numerical thread each, at most two kernel builds per worker and a one-hour cooperative ceiling. Expected runtime is about 5–20 minutes. Separately precharge at most 620 bounded linear operations per fit, 100 iterations each. Freeze source, environment, identities, selection, allowances and software qualification before launch. Preserve every failure and close all unused work at termination.
Broader gate and interpretation
Only if all four regression fits, replay checks and required start comparisons pass, prepare a separate replication of the earlier 48-case optimizer qualification: its same exposed development seeds, two input laws, all three candidates, both processing settings and two starts. Retain its 200-evaluation per-fit limit and derive a new finite source-bound allowance; do not reuse its closed budget. Freeze that phase separately before execution.
A failed gate blocks its dependent phase. Neither stage reopens a calibration campaign, estimates a threshold or uses protected aircraft observations. Native optimizer qualification is a prerequisite for proposing another calibration release, not permission to restart one automatically.
Outcome: all four regressions return, none qualifies. Original bindings replay exactly. Final cost stability fails in distant midpoint-anchor checks, while stationarity and physical-change checks pass. The broader gate stays blocked. The local-anchor plan defines a separate software refinement and qualification; this allowance is closed.