# Fast Feature V1 Experiment

- Episodes: 5
- Baseline extraction: 361.90s
- Fast extraction: 246.47s
- Wall-time reduction: 31.9%
- Baseline raw visual GT recall: 16/17
- Fast raw visual GT recall: 17/17

| Episode | Baseline | Fast | Faster | Face valid Δ | Face IoU | Baseline→Fast | Fast→Baseline | GT baseline | GT fast |
|---|---:|---:|---:|---:|---:|---:|---:|---:|---:|
| P24/EP20 | 73.24s | 50.99s | 30.4% | +2.8pp | 0.9652 | 100.0% | 98.2% | 4/5 | 5/5 |
| P24/EP21 | 70.67s | 48.37s | 31.6% | +4.8pp | 0.9742 | 99.4% | 98.8% | 1/1 | 1/1 |
| P24/EP23 | 69.86s | 47.83s | 31.5% | +3.7pp | 0.9698 | 100.0% | 97.3% | 3/3 | 3/3 |
| P24/EP25 | 80.80s | 58.08s | 28.1% | +5.9pp | 0.9744 | 99.5% | 97.3% | 1/1 | 1/1 |
| P26/EP16 | 67.33s | 41.20s | 38.8% | +4.5pp | 0.9732 | 98.1% | 96.2% | 7/7 | 7/7 |

## Family overlap

| Family | Baseline→Fast | Fast→Baseline |
|---|---:|---:|
| black | 5/5 (100.0%) | 5/5 (100.0%) |
| blur | 212/213 (99.5%) | 214/218 (98.2%) |
| deformation | 280/284 (98.6%) | 270/287 (94.1%) |
| flicker | 246/246 (100.0%) | 246/246 (100.0%) |
| judder | 115/115 (100.0%) | 115/115 (100.0%) |

## Interpretation

- Fast extraction preserves the full per-frame timeline; no global FPS reduction is used.
- Black, flicker, and judder candidates are unchanged within the 0.5s comparison tolerance.
- Differences are concentrated in marginal deformation and blur candidates; raw visual GT recall did not decrease in this five-episode sample.
- The extra EP20 GT hit is tolerance-edge coverage and should not be interpreted as proof that fast mode improves production recall.

Candidate overlap uses same-family intervals with 0.5s tolerance. GT recall is raw detector coverage with 1.0s tolerance and excludes lip-sync/out-of-scope rows. This experiment does not measure post-VLM ranking precision.