Compiled-monitor replay on Stress benchmark
54True Positives (TP)Gold
Evaluation Results
| Method | Links | ||||||
|---|---|---|---|---|---|---|---|
| GoldType=Oracle2026.06 | 54 | 16 | 4 | 93.1 | 77.1 | 84.4 | |
| LoRAModel=Qwen3-8B2026.06 | 41 | 29 | 7 | 85.4 | 58.6 | 69.5 | |
| LabGuard-Hybrid2026.06 | 41 | 29 | 7 | 85.4 | 58.6 | 69.5 | |
| RegexType=Rule-based2026.06 | 34 | 36 | 9 | 79.1 | 48.6 | 60.2 | |
| No SafetySafety Monitoring=None2026.06 | 7 | 63 | 1 | 87.5 | 10 | 17.9 |