Staged training: generated software controls
6/6 cases pass; elapsed 94.31 seconds.
| Candidate | Processing | Saved outcome | Future widths N / E / height |
|---|---|---|---|
| Specified-field globe | Zero removal | Pass | 0.0051112 / 0.0051435 / 0.0047282 m |
| Specified-field globe | Profiled removal | Pass | 0.0051806 / 0.0052128 / 0.0047976 m |
| Zero-field globe | Zero removal | Pass | 0.0051088 / 0.005126 / 0.0047261 m |
| Zero-field globe | Profiled removal | Pass | 0.0051829 / 0.0051999 / 0.0048002 m |
| Flat disc | Zero removal | Pass | 0.0050819 / 0.0044582 / 0.0047193 m |
| Flat disc | Profiled removal | Pass | 0.005157 / 0.0044812 / 0.0047944 m |
Charged 60/60 native queries, 30/30 bounds and 8640/8,640 logical bound steps.
The two training stages retain every earlier inequality and all 27 physical directions. Passing cases include independent exact prediction-endpoint checks and a feasible known witness examined after saving the prediction. Native warnings and all reservations are preserved.
These are 32-packet examples with a 0.13-second prediction and a supplied 1 mm observation allowance. They check the staged software path, not recording-length usefulness, an observational error law, whole-program floating-point enclosure or independent physical truth. No aircraft or protected ESA measurements are used.