Fresh prediction calibration: why it stopped
The run stopped at a numerical check. It established no uncertainty threshold, and independent validation has not started. The original 53-stretch aircraft comparison is unchanged.
The corrected arithmetic had passed 48 development fits. This larger test used fresh simulated measurements to check whether prediction errors could support a reliable uncertainty threshold. Every physical model fitted its own generated truth: these examples are not votes for globe or flat geometry.
What completed
| Item | Verified outcome |
|---|---|
| Software qualification before launch | 126 source-bound checks passed. |
| Returned starts | 909 of 7,176 planned, in 9.58 hours. |
| Numerical outcomes | 904 passed; two failed; three stopped cooperatively. |
| Position-error screens | All 906 completed forecasts passed. |
| Complete noise families with available scores | 37 of 299 planned. Related settings and repeated starts are not independent trials. |
| Calibration threshold | Unavailable. The incomplete cohort cannot supply the planned bound. |
| Independent validation | Not launched. Its separate seed cohort remains unused. |
All workers exited and released their locks. The remaining allowance is closed; the run will not be extended or restarted.
The specific failure
Both prescribed starts failed for the same noise family: seed 281474977010693, rate-sample input errors, specified-field globe, with upstream field removal allowed. They stopped after 43 and 41 evaluations, below the 200-evaluation limit. Both reported no positive predicted decrease.
Their position predictions were accurate under the fixed screens: later horizontal RMS was about 2.58 m, and height RMS was about 0.31 m. The numerical acceptance check asks a different question: does one fresh bounded correction leave the cost, predicted positions and stationarity stable?
The stationarity and position-change checks passed. The relative cost changed by about 6.27 × 10⁻⁸, exceeding the unchanged 10⁻⁸ limit by about sixfold. Maximum position change was under 0.20 mm. These remain failed fits; small position changes do not override the frozen cost requirement.
Once both starts failed, a required family score was unavailable. The frozen rule therefore stopped the other workers cooperatively. Three interrupted starts remain recorded, alongside every unexecuted planned row.
What follows
The saved fits identify a cost-stability failure; they do not establish its underlying cause or justify relaxing the limit. Any revised numerical method needs separate qualification and a reviewed finite plan before another campaign. No fits or native calculations were run to produce this report.
A publication-path repair passes 106 software checks, including a regression that copies the actual frozen plan through the shared packager. Only the reporter and its test changed after execution; the 242 execution-source snapshots, numerical environment, scientific settings and saved outcomes remain preserved.
This test concerns two-minute generated trajectories and declared error laws. It does not establish actual aircraft error coverage, calibrated significance for the 53 stretches, or readiness to open protected ESA confirmation groups.
Machine-checked report · All planned starts · Frozen plan · Earlier calibration outcome · Next analyses · Reproducibility.