Recovery of cost-only numerical failures

Completed: 83/86 fits passed in 891 native sweeps and about four hours. Both controls and all twelve repeats passed. Nine more stretches have complete model comparisons; three remain unfinished after exhausting their primary allowances. The combined findings are 41 globe preferences, zero flat and 12 unresolved among the original 53 selected stretches. Results preserve every outcome. The paired precision qualification passed all 16 fits, including both controls and both repeats, with unchanged physical assumptions and numerical tolerances.

This separate study applied that qualified arithmetic to all twelve remaining stretches whose failed stopping checks concern only relative objective change. Two have complete primary fits but an unfinished repeat. Selection uses saved failure types, never a preferred model or a low objective value. The other nine stretches have stationarity or position-stability failures and are excluded from this particular recovery. The inventory records every remaining stretch and its inclusion or exclusion reason.

Each selected stretch received all three physical models under both processing assumptions, followed by its original predetermined perturbed-start repeat. Two previously completed recordings provide fresh controls. Every start uses the frozen primary endpoint; repeats never replace primaries. The whole selected interval, measurements, bounds, weights, integration clock and stopping limits remain unchanged. Earlier evidence and failures remain preserved.

The finite plan caps the study at 86 starts and 5,532 native sweeps, using one worker with four numerical threads. There are no retries or automatic budget increases. Journals charge work before execution; exclusive locks and hashed atomic outputs support safe recovery after interruption. A completed study cannot restart. Evidence packaging is complete. The freeze is closed; unused allowances cannot be reused.

A numerical preference requires all six primary fits, its scheduled repeat and both controls to pass. Partial comparisons remain unresolved. This study does not calibrate statistical significance, prove a global optimum or identify the physical cause of the background rotation field.