Check the saved bounded linear calculation
Question
Can the current bounded linear calculation return a worse objective than a feasible zero step while reporting success? The closed local-timing calibration contains two such verification records. This review must distinguish saved evidence from a reconstruction using a different linearization.
Fixed selection and limits
Use the two failed starts for seed 281474977210667, increment input, specified-field globe, no upstream-removal parameter. Include both prescribed starts of the closest earlier completed seed with the same configuration and two numerical passes. Select that control using training/completion metadata, without later prediction errors. Bind the parent manifest, summary and fit hashes.
Use only each saved final residual, full 26-column Jacobian, fitted point and the declared timing cell containing that exact point. Do not shift the anchor or reconstruct unsaved residuals/Jacobians. This is a baseline-linearization check, not an exact replay of the earlier verification at a different anchor. Report the original saved verification metrics separately.
Freeze four records, at most four bounded linear solves, one numerical thread and a 600-second cooperative ceiling. Charge before solving. No native calls, nonlinear fits, new noise draws, aircraft observations or protected data. Earlier unused allowances remain closed. Keep all columns and original bounds.
Evidence and next gate
Recompute proposal residuals and costs with independent scalar arithmetic. Report feasibility, the zero-step benchmark, cost change and the solver's success flag separately. A completed diagnosis does not qualify a solver or change any historical fit label. Preserve failures and missing outcomes.
After source-bound software checks, use the shared study lifecycle, verify the closed results and publish a short report and CSV. Any proposed solver change then needs independent known-answer tests before a separately frozen trial. Calibration and validation cannot restart from this review.