Fresh-draw fitting and prediction pilot

This is a generated-data diagnostic of the complete fitting procedure. It does not open another aircraft interval or a reserved ESA group. The saved-score study and every earlier failed numerical or accuracy check remain unchanged.

Fixed question and matrix

Repeat the existing two-minute noisy training/prediction procedure with two preassigned fresh seeds: 26102610 and 26102611. Their role is method_diagnostic_pilot; they cannot later estimate a calibration threshold or serve as independent validation. Software seed checks use different seeds.

For each seed retain both input-error formats, all three generating/fitted candidates, both processing assumptions and both declared starts. This makes 48 starts. Each candidate is fitted to data generated under that same candidate; this is not a three-way Earth-model comparison.

Preserve 300 Hz sampling, 120 seconds, 119 receiver epochs, the first 79 epochs for training, the original calibration bounds and the 200-evaluation limit per start. Keep the original 10 m horizontal / 30 m height descriptive screen. Estimate calibration anew from earlier observations; never supply true initial state, true calibration or later positions to fitting or weights. Save forecasts before later truth. Use training cost and declared order for primary selection.

Qualification before launch

The seed interface must pass its frozen-generator compatibility checks. Extract the existing structured_noisy_study.evaluate execution body into one shared case runner, preserving its prepayment, construction journal, native factory and failure handling. Keep the original evaluator as its strict caller. The new small lll.independent_noisy_study definition calls that runner and the shared prediction scorer; it must not copy a worker or navigation library.

Qualify lifecycle binding, immutable seed identity and retained failed outcomes. Reuse the twelve short native protocol controls at the separate qualification seed 26102602, including both formats and all model/processing cases. Add the score check to the real saved forecast path. The short native qualification keeps its existing 1,000-call / 10,000-step / 30,000-query-visit limits per case, eight fit evaluations, two shared libraries and five-minute cooperative ceiling. Its purpose is correct execution and honest labels, not a positive confidence claim.

Separate finite allowance

Use two workers with one numerical thread each, reserving two threads while the existing full test job owns its allocation. Expect roughly two hours at the earlier measured throughput; the new cooperative ceiling is four hours. Do not terminate a process to enforce it. No retry or reuse of another study's unused allowance is permitted.

Each worker may build the existing noise bridge and wider-tangent library once. An operator remains capped at 64 MiB. Native counts and visits retain the existing conservative accounting convention. These are new limits for this fresh question, not an expansion of the completed study.

Report every outcome

Retain all source/environment bindings, seeds, raw-input hashes, construction charges, fit records, error covariances and later error curves. Report numerical passes and failures separately from physical accuracy and normalized scores. Select primaries without later errors; retain unavailable groups.

Compute each complete-block coordinate error/noise score with the shared scorer. The covariance still treats fitted calibration as known. Two new seeds can expose instability and test the pipeline; they cannot calibrate a rare tail, establish predictive coverage or validate an Earth-model decision. Further calibration and validation draws require separate disjoint identities and freezes.

Independent draws · Roadmap.