Policy Violation Enforcement on MedAgentBench (Adversarial)
0Unsafe Rate (Hong et al.)Hong et al. (Manual Guardrails)
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| Hong et al. (Manual Guardrails)Condition=Guardrail2026.06 | 0 | — | |
| Hong et al. (Manual Guardrails)Condition=Baseline2026.06 | 62 | — | |
| Hong et al. (Manual Guardrails)Condition=Raw2026.06 | 78 | — | |
| Verification Sandwich (Cedar)Condition=Raw2026.06 | — | 72 | |
| Verification Sandwich (Cedar)Condition=Baseline2026.06 | — | 82 | |
| Verification Sandwich (Cedar)Condition=Guardrail2026.06 | — | 85.7 |