Policy Violation Detection on PhantomPolicy safe-control original
0.0333Violation RateGPT-5.4
Evaluation Results
| Method | Links | |
|---|---|---|
| GPT-5.4Condition=Execution-oriented baseline, Labels=Human-reviewed trace labels2026.04 | 0.0333 | |
| GPT-5 miniCondition=Execution-oriented baseline, Labels=Human-reviewed trace labels2026.04 | 0.0333 | |
| GPT-5.4 nanoCondition=Execution-oriented baseline, Labels=Human-reviewed trace labels2026.04 | 0.05 | |
| Claude Opus 4.6Condition=Execution-oriented baseline, Labels=Human-reviewed trace labels2026.04 | 0.05 | |
| Claude Sonnet 4.6Condition=Execution-oriented baseline, Labels=Human-reviewed trace labels2026.04 | 0.1333 |