High-precision candidate selection on document-level 60/20/20 (test)
2,743Number of SelectionsOpenAI ∪ Claude (closed ≥ 1)
Evaluation results
| Method | Links | ||||||
|---|---|---|---|---|---|---|---|
| OpenAI ∪ Claude (closed ≥ 1)Threshold Strategy=no threshold tuning2026.05 | 2,743 | 743 | 72.9 | 94.2 | 83.2 | 88.3 | |
| Agreement ≥ 2Threshold Strategy=no threshold tuning2026.05 | 2,587 | 797 | 69.2 | 84.3 | 74.5 | 88.3 | |
| BioConCal (GBT)Threshold Strategy=Validation-frozen P95 threshold2026.05 | 1,340 | 82 | 93.9 | 59.2 | 52.3 | 88.3 | |
| Agreement ≥ 4Threshold Strategy=no threshold tuning2026.05 | 1,275 | 260 | 79.6 | 47.8 | 42.2 | 88.3 | |
| BioConCal (logistic)Threshold Strategy=Validation-frozen P95 threshold2026.05 | 1,229 | 85 | 93.1 | 53.9 | 47.6 | 88.3 | |
| Agreement ≥ 6Threshold Strategy=no threshold tuning2026.05 | 715 | 93 | 87 | 29.3 | 25.9 | 88.3 | |
| Unanimous (k = 8)Threshold Strategy=no threshold tuning, k=82026.05 | 293 | 14 | 95.2 | 13.1 | 11.6 | 88.3 | |
| Raw agreement / 8Threshold Strategy=Validation-frozen P95 threshold2026.05 | 293 | 14 | 95.2 | 13.1 | 11.6 | 88.3 | |
| Isotonic on agreementThreshold Strategy=Validation-frozen P95 threshold2026.05 | 293 | 14 | 95.2 | 13.1 | 11.6 | 88.3 |