Native timing verification does not yet qualify
The four original-start regressions finished, but none passes the complete numerical gate. All four reproduce the original input arrays, training observations and error operators exactly. Their raw forecasts meet the declared position-error screens; the final cost-stability check fails.
The run takes 10.1 minutes with four one-thread workers. It uses 238,386 native operations, including input/error propagation, and 126 bounded linear operations. The fits themselves use 25, 25, 61 and 59 evaluations, all below their unchanged 200-evaluation ceiling. All unused work allowances are closed.
| Existing example | Earlier fit passed | New fit passed | Later horizontal RMS | Later height RMS |
|---|---|---|---|---|
| Passing control, first start | Yes | No | 4.494 m | 0.752 m |
| Passing control, second start | Yes | No | 4.494 m | 0.752 m |
| Earlier numerical failure, first start | No | No | 2.576 m | 0.307 m |
| Earlier numerical failure, second start | No | No | 2.576 m | 0.307 m |
These are exposed generated-data cases, not new aircraft measurements. The earlier results keep their original labels. This run supplies no uncertainty threshold and does not change the 53-stretch findings.
What the failure shows
All six failed cell checks pass stationarity and physical position-change limits. Their relative cost changes range from about 1.07 × 10⁻⁸ to 1.88 × 10⁻⁸, above the unchanged 10⁻⁸ limit. The largest RMS position change is about 0.12 millimetres, below the 1-millimetre limit.
The failed checks use timing-cell midpoint anchors about 1.67 milliseconds from the fitted offset. The earlier successful fixed-proposal study used anchors only 0.1–10 microseconds from the boundary. That difference suggests the new check is linearizing too far from the point it is trying to verify.
Independent software control
A known curved objective, 1 + 10⁻⁶ × (x + x³)², has its exact minimum at
x = 0. The unchanged verifier rejects that minimum when cell midpoints are
used: its cost change is about 2.12 × 10⁻⁸. Projecting the fitted timing value
into each cell instead makes the check pass, with zero cost change. Both
versions pass the original stationarity and physical-change screens.
This control uses ten analytical residual evaluations, four bounded linear operations, no native propagation and no new fits. Its source-bound tests and source map are retained with the publication. It demonstrates a false-failure mechanism and supports a specific next software test; it does not establish that a revised native optimizer will qualify.
The local-anchor plan keeps every timing cell and nuisance parameter, with the same thresholds. It requires an expanded analytical qualification before a separately frozen native regression. The broader 48-case run and calibration remain blocked.
All fits · Every failed check · Start comparisons · Frozen native plan · Reproduction record · Next analyses.