Can fixed solver scaling finish the prediction comparison?

Completed: five of six fits converged; comparison qualification did not pass. The flat-disc fit with upstream removal allowed exhausted its 200-evaluation limit. The completed prediction trial had five converged fits and one unfinished flat-disc fit. This separate comparison preserves the flight, models, calibration bounds, timing and training/prediction split. It changes how the numerical search scales parameter steps, while recording more diagnostic information.

Both dense solver settings passed 24/24 known-optimum controls. The implemented candidate passed its focused tests. These are software checks; the matched airborne trial did not recover the unfinished comparison. See the analytic screen and saved prediction audit.

The plan specifies six new fits, 200 evaluations per fit and no retries, with two numerical threads. Estimated runtime is roughly 55–65 minutes if every fit uses its allowance. The worker preserves every result and publishes its evidence when finished. A separate metadata-only watcher then verifies the numerical freeze and builds the matched report automatically; it performs no fits. Runtime progress and report completion are recorded separately in the study's data directory. The original unfinished fit stays preserved, and this study does not authorize any archive-wide rerun or change to the scientific evidence gate.

The matched results show which fits converged, whether the five previously completed objectives were preserved, and how prediction errors changed. The six starts used 749 evaluations and about 33.4 minutes of fitting and diagnostics. The no-removal prediction comparison favors globe; the full two-case comparison remains unresolved. More completed fits alone do not establish physical calibration or statistical significance. No retries or replication starts are authorized.