Reconstruction allowances dominate the prediction widths
All six saved cases pass all 18 exact width checks. The nonlinear reconstruction allowance accounts for 76.7–90.8% of the complete prediction widths. This gives a more useful denominator than the earlier audit, which compared position reconstruction only with the unshared allowance.
These are conservative ranges from generated software tests. They are not measured aircraft position errors. No model was fitted or trajectory reintegrated.
| Candidate | Removal assumption | North width, km | East width, km | Height width, km | Reconstruction share, N / E / height |
|---|---|---|---|---|---|
| Globe, specified field | None | 2026.6 | 2413.3 | 1462.7 | 87.5% / 88.7% / 90.7% |
| Globe, specified field | Profiled | 2024.5 | 2418.7 | 1461.0 | 87.4% / 88.5% / 90.5% |
| Globe, zero field | None | 1704.1 | 1995.6 | 1258.8 | 87.7% / 88.7% / 90.8% |
| Globe, zero field | Profiled | 1699.8 | 2001.9 | 1253.0 | 87.7% / 88.6% / 90.7% |
| Flat disc | None | 657.3 | 109.3 | 312.8 | 77.4% / 78.0% / 77.4% |
| Flat disc | Profiled | 687.5 | 114.9 | 339.3 | 76.7% / 76.9% / 77.3% |
The remaining width is mainly the affine range over permitted calibration parameters. Arithmetic and outward-reporting increments are also retained and checked exactly; the largest width-reporting increment is less than 0.000000001 m. Improving floating-point subtraction cannot resolve these ranges.
The next method must tighten the nonlinear reconstruction allowance while preserving the physical bounds and earlier observation constraints. The static coupling comparison tightens a reference-acceleration allowance by 3.7–7.1%, but has not recomputed these trajectory ranges. Further diagnosis should target their larger time-varying force and mounting-transform contributions before repeating the full-history integration.
The audit took 1.64 seconds, with no native calls, new support solves, aircraft reads or optimizer starts. The first audit's rounding-check failures are preserved separately. Original prediction failures and all 53 main model comparisons remain unchanged.
Method and limits · First audit · Shared arithmetic · Prediction requirements · Reproducibility.