# bench_score

Per-(prediction-repo-sha × model) IoU-24 scores for the BenchCAD/cad_bench_200
benchmark. Updated by `eval/remote_iou24_service.py` (see cadrille repo).

Layout:
```
per_case/{sha[:8]}/{model}.parquet   # one row per stem
summary.parquet                       # rolled-up by (sha, model)
```

Per-case parquet schema:
  sha_short, model, stem, code_hash, iou, iou_24, rot_idx, exec_ok,
  reason, latency_s, scored_at

Summary schema:
  sha_short, model, n_total, n_exec_ok, exec_rate,
  mean_iou_24, median_iou_24, pass_iou_24_0.5, pass_iou_24_0.7,
  updated_at