# Results

- cases: **622**
- prototypes: **989** (tooth-aware=True, assign threshold=0.45)
- out-of-fold mAP: **0.0738** (folds: 0.0883, 0.0667, 0.0694, 0.0805, 0.0640)

## Out-of-fold score

| metric | value |
|---|---|
| final (0.8 clinical + 0.2 captioning) | **0.3544** |
| clinical F1 (surrogate) | 0.3861 |
| logical precision | 0.4910 |
| logical recall | 0.3181 |
| BLEU-4 (grader-exact) | 0.1493 |
| METEOR (grader-exact) | 0.3064 |

## Ablation

| variant | final | clinical | BLEU-4 | METEOR |
|---|---|---|---|---|
| corpus prior (no imaging) | 0.3575 | 0.3937 | 0.1170 | 0.3082 |
| linear, 122 features | 0.3544 | 0.3861 | 0.1493 | 0.3064 |
| fine-tuned encoder (29M params) | 0.3403 | 0.3752 | 0.1331 | 0.2684 |

## Caveats

- Clinical scores are the repository's offline RadFact surrogate unless 'evaluation_radfact_lite' is populated. The surrogate ranks decoder variants; it is not the challenge metric.
- BLEU-4 and METEOR are exact reimplementations of the grader's local implementations and match NLTK to machine precision.
- Scores are out-of-fold on the public training release, not the hidden 50-case test set, and the folds are stratified rather than leave-one-centre-out, so they measure in-domain performance.