Policy Enforcement on PhantomPolicy (human-reviewed trace labels)
40.7Risky-case Violation RatePolicy-in-prompt
Evaluation Results
| Method | Links | ||||||||
|---|---|---|---|---|---|---|---|---|---|
| Policy-in-promptStrategy=Policy-in-prompt2026.04 | 40.7 | — | — | — | — | 5.7 | — | — | |
| BaselineStrategy=No enforcement2026.04 | 95.3 | — | — | — | — | — | — | — | |
| Content DLPStrategy=Content DLP2026.04 | — | 68.83 | 56.61 | 96.06 | 40.13 | — | — | — | |
| SentinelStrategy=Sentinel2026.04 | — | 92.99 | 92.71 | — | — | — | 5 | 37 |