Bounded archive comparison: preparation requirements

This plan permits preparation, not observed fits, native predictions, retries or a corpus rerun. Preserve the original findings and all failed-fit labels.

1. Choose development records and splits before fitting. Start with the already used October 5 control and the predeclared September 24 recording. They provide retrospective development checks, not unseen confirmation. Keep overlapping streams and repeated calibration grids grouped. Preserve full-record drift anchors.
2. Declare one finite covariance list. Start with one correlated component and a positive white floor. Justify cross-axis terms and time scales independently of each competing model's preferred residuals. Estimate on designated training data only, seal the chosen specification, and apply it unchanged to all three models and processing cases. Preserve ambiguity between covariance candidates.
3. Prepare the native prediction/derivative adapter. Validate the full nonlinear derivative matrix and receiver association on existing independent geographic controls. Check rank using every permitted calibration direction, including parameters at bounds. Do not treat a singular inverse or fixed-bound covariance as established precision.
4. Specify uncertainty treatment before seeing new outcomes. Interior local linear prediction is a candidate only after noise-weighted nonlinear probes cover the relevant directions and scales. At bounds, use feasible nonlinear sensitivity points or genuinely profiled fits, labeling finite-grid extrema appropriately. Their cost allowance is descriptive unless independently calibrated. Calibrated predictive intervals need a justified bounded estimator distribution and covariance-parameter uncertainty; neither is supplied by these controls.
5. Seal identities and cost before launch. Separate covariance estimation, training fits and uncertainty profiles in the start/evaluation ledger. Count every native derivative/prediction pass, failed start and profile reoptimization. Declare finite limits, no adaptive retries and at most two total numerical threads. Ask for the resulting observed-data allowance only after the adapter and plan are reviewable. Never reuse unused allowances from completed studies.
6. Compare physical predictions and adequacy first. Save held-out errors and covariance sensitivities for all candidates. Report globe/disc before a background- field comparison, retain numerical abstentions, and distinguish fitted model preference from statistical detection and the field's physical cause.

The immediate next implementation is the control-tested native uncertainty adapter and a concrete, costed freeze. Full-archive rollout requires a successful bounded comparison and separate authorization.