We can trust the confidence number. Auto-accept the confident lines; route only the shaky ones to a human or the big-model backup, instead of reviewing everything. It also gives customers a real confidence badge.
| Confidence group | Lines | Land on a real stripe |
|---|---|---|
| Q1 (lowest) | 257 | 27% |
| Q2 (mid) | 257 | 97% |
| Q3 (mid) | 257 | 100% |
| Q4 (mid) | 257 | 100% |
| Q5 (highest) | 258 | 100% |
Every line split into five equal groups by the model's own confidence, then checked against real human stripes.
Source: 332-photo holdout, frozen 2026-08-25. 1,286 detections, quintile buckets.