More training positions: a generated-data diagnostic
6/6 software diagnostics pass; 6/6 completed.
| Model | Processing | Five epochs: N / E / height | 292 epochs: N / E / height | Linear target met? |
|---|---|---|---|---|
| Specified-field globe | Zero removal | 420.263 / 747.100 / 39.927 m | 292.214 / 283.105 / 21.768 m | No |
| Specified-field globe | Profiled removal | 445.141 / 1068.424 / 54.499 m | 295.808 / 416.223 / 25.690 m | No |
| Zero-field globe | Zero removal | 420.160 / 748.501 / 49.232 m | 291.454 / 285.634 / 24.983 m | No |
| Zero-field globe | Profiled removal | 460.684 / 1071.583 / 68.364 m | 291.491 / 421.390 / 30.674 m | No |
| Flat disc | Zero removal | 447.043 / 150.881 / 70.007 m | 311.288 / 123.450 / 54.908 m | No |
| Flat disc | Profiled removal | 515.579 / 245.023 / 111.322 m | 337.358 / 154.925 / 70.726 m | No |
These are exactly checked outer bounds for the saved linearized prediction, not nonlinear prediction ranges or aircraft confidence intervals. All 27 physical columns and their original broad domain remain. Shared errors and nonlinear terms are omitted in this diagnostic.
The dense set includes every sparse training epoch. Native batch/endpoints agree exactly, and packet-boundary counters are preserved and checked. The target remains 20 / 20 / 50 m. A numerical pass and a useful linear width are different outcomes. Neither establishes useful full-history nonlinear bounds, an observation-error law or scientific significance.