Safety violation detection on SafetyDrift traces (test)
94.7Detection RateSAFETYDRIFT
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| SAFETYDRIFTms/step=<0.01, Model Order=1st order2026.03 | 94.7 | 11.8 | |
| LLM Judge (3-step)ms/step=646, Model=Claude Haiku, Context Window=3-step2026.03 | 57.9 | 47.1 | |
| LLM Judgems/step=588, Model=Claude Haiku2026.03 | 52.6 | 38.2 | |
| Keywordms/step=<0.012026.03 | 44.7 | 55.9 | |
| No Monitorms/step=0.02026.03 | 0 | 0 |