This was the cheap test: if the model truly never learned faint stripes exist, no amount of re-running it differently should conjure them — and that's roughly what happened. The weak recovery, even with 5 extra passes, is itself evidence the miss is a LABEL problem (the model was never taught these stripes), not something more computation alone can fix. Points toward the label-audit work (test 0.2) as the real fix.
| Pass | Recall | Precision |
|---|---|---|
| Baseline (rules stage) | 94.8% | 98.1% |
| + TTA, 5 passes merged (rules stage) | 98.0% | 90.0% |
| Ledger | Count |
|---|---|
| Stripes recovered that we were missing | 35 |
| New false lines added | 132 (66 duplicates of a real stripe + 91 nowhere near a real stripe) |
| Real stripes lost | 0 |
We ran our existing model 5 different ways on each photo (different zoom, a zoomed-in bottom crop, a lower confidence bar) and merged what came back. No retraining, no new model.
Source: 332-photo holdout, frozen 2026-08-25. 5 TTA passes (base/960/1280/bottom-crop/low-conf), custom polyline fusion.