NDA non-compliance clause detection on Internal NDA 40-policy NDA-2
100AccuracyPolicyGuard
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| PolicyGuardModel=GPT-4.1, Protocol=zero-shot2026.06 | 100 | 100 | |
| PolicyGuardBackbone=GPT-4.1, Zero-shot protocol=true2026.06 | 95.3 | 85 | |
| Claude CoworkModel=Opus-4.6, Guidelines=raw policy guidelines, Protocol=zero-shot2026.06 | 92.5 | 84.2 | |
| PolicyGuardModel=Sonnet-4.5, Protocol=zero-shot2026.06 | 87.5 | 76.2 | |
| Majority ClassBackbone=GPT-4.1, Zero-shot protocol=true2026.06 | 86.3 | — | |
| Claude CoworkModel=Sonnet-4.6, Guidelines=raw policy guidelines, Protocol=zero-shot2026.06 | 80 | 57.1 | |
| Direct PromptingBackbone=GPT-4.1, Zero-shot protocol=true2026.06 | 73.2 | 45.7 | |
| Legal SyllogismBackbone=GPT-4.1, Zero-shot protocol=true2026.06 | 70.2 | 45.5 | |
| Chain-of-ThoughtBackbone=GPT-4.1, Zero-shot protocol=true2026.06 | 69.8 | 43.2 | |
| Legal ReasoningBackbone=GPT-4.1, Zero-shot protocol=true2026.06 | 69.8 | 46.4 |