Full matrix check of the corrected optimizer
Question: does the corrected bounded linear calculation work across both input formats, all three generating models and both processing settings? The six-case regression passes, including the latest calibration failure at its unchanged fitted cost. This broader test must pass before another calibration release is considered.
Inputs and method
Reuse the 48 exposed development cases from the original matrix: seeds 281474976810725 and 281474976810726; point-rate and increment inputs; the specified-field globe, zero-field globe and stationary disc; two processing settings; and the same two declared starts. Each candidate fits its own generated truth. These are software controls, not 48 new model votes.
Keep the original observations, error law, physical bounds, starting points, all 26 or 27 nuisance columns and every acceptance limit. Preserve four exact original input/training/error-operator replays and the original failure labels. The corrected calculation starts each bounded linear solve at its feasible zero-step anchor and rejects worsening answers. No scientific threshold changes.
The prerequisite is the independently audited six-fit publication, bound by
manifest f2c63e792616bcaf9909a61da562339314935f552f3d1df73fbf7395e41bbd6b
and summary b9031118b5a3d5fef32e6907e33dcc339c4f15698dead97bacddbad9aaa02adf.
All seven numerical source hashes and the runtime environment must match.
Only the documented reporting repair differs from that execution freeze.
Use the shared study worker, counted solver and independent saved-result auditor;
the new module contains only this study's policy and metadata.
Finite allowance
Freeze sources, environment, task table, selection and software qualification before execution. Use four workers with one numerical thread each. The ceiling is 48 starts, 200 evaluations and 620 bounded linear operations per start, 100 iterations per operation, with no retries. Maximum native prepayments are 2,877,792 calls and 18,072,612,144 packet visits. The cooperative wall ceiling is two hours; the preceding matrix took about 66 minutes. Charge before work and close all unused allowance at completion. No closed study is resumed.
Acceptance and next step
Require all 48 numerical and physical forecast checks, all 24 paired-start comparisons, and all four original replays to pass. Keep the existing training cost/position repeat limits and the separate later horizontal/height accuracy screens. Publish every result, including failures, without refitting during audit.
A pass qualifies this generated-data matrix. It does not establish calibrated prediction intervals or confidence for aircraft results. Any new calibration requires disjoint cohorts and a separately reviewed finite allowance; all previous calibration allowances stay closed. The original 53 aircraft results and protected ESA confirmation groups remain unchanged.