PPC kinematic integrity six-fold LOO¶
Reference truth is used only for offline labels and candidate scoring. Runtime inputs are consecutive POS positions and timestamps.
| Held out | Selected jump / acceleration / hold | Train >10m caught | Train correct harm | Held >10m caught | Held correct harm |
|---|---|---|---|---|---|
| tokyo_run1 | 12 m / 200 m/s2 / 3 | 20 | 12 | 6 | 3 |
| tokyo_run2 | 12 m / 200 m/s2 / 3 | 23 | 12 | 3 | 3 |
| tokyo_run3 | 12 m / 200 m/s2 / 5 | 22 | 14 | 8 | 13 |
| nagoya_run1 | 12 m / 200 m/s2 / 3 | 26 | 15 | 0 | 0 |
| nagoya_run2 | 12 m / 200 m/s2 / 3 | 26 | 13 | 0 | 2 |
| nagoya_run3 | 12 m / 200 m/s2 / 3 | 17 | 15 | 9 | 0 |
Production: 12 m / 200 m/s2 / 3 epochs.
It catches 26 >10 m wrong FIX and 30 total wrong FIX while harming 15 correct FIX epochs.
Advanced extension LOO¶
The primary production gate is fixed. Each fold selects only plateau continuation and the telemetry-backed secondary trigger from the other five runs.
| Held out | Plateau / max age / secondary | Train added >10m | Train added correct harm | Held added >10m | Held added correct harm |
|---|---|---|---|---|---|
| tokyo_run1 | 0.1 m / 8 / prefit>5, ratio<=15, outliers>=10, nsat<=15 | 11 | 8 | 2 | 5 |
| tokyo_run2 | 0.1 m / 8 / prefit>5, ratio<=15, outliers>=10, nsat<=15 | 9 | 8 | 4 | 5 |
| tokyo_run3 | 0.1 m / 8 / prefit>5, ratio<=15, outliers>=10, nsat<=15 | 13 | 10 | 0 | 3 |
| nagoya_run1 | 0.1 m / 8 / prefit>5, ratio<=15, outliers>=10, nsat<=15 | 13 | 13 | 0 | 0 |
| nagoya_run2 | 0.1 m / 8 / prefit>5, ratio<=8, outliers>=10, nsat<=12 | 9 | 10 | 0 | 0 |
| nagoya_run3 | 0.05 m / 8 / prefit>5, ratio<=15, outliers>=10, nsat<=15 | 10 | 13 | 0 | 0 |
Advanced production catches 39 >10 m and 43 total wrong FIX while harming 25 correct FIX epochs. The extension alone adds 13 / 13 catches at a cost of 10 correct epochs.