Misalignment Detection on Workbench n=62 (general)
0.3Error Rate (ER)ProvenanceGuard
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| ProvenanceGuardModel=Gemini 3 Flash2026.05 | 0.3 | — | — | |
| ProvePromptModel=Gemini 3 Flash2026.05 | 1.6 | — | — | |
| ProvePromptModel=GPT-52026.05 | 1.9 | — | — | |
| DirectPromptModel=GPT-52026.05 | 2.3 | — | — | |
| ProvenanceGuardModel=GPT-5 Mini2026.05 | 2.9 | — | — | |
| DirectPromptModel=Gemini 3 Flash2026.05 | 3.5 | — | — | |
| ProvenanceGuardModel=Claude Haiku 4.52026.05 | 5.6 | — | — | |
| ProvenanceGuardModel=GPT-52026.05 | 7.4 | — | — | |
| ProvePromptModel=Gemini 2.5 Flash2026.05 | 9.7 | — | — | |
| ProvePromptModel=GPT-5 Mini2026.05 | 10.3 | — | — | |
| ProvePromptModel=Claude Sonnet 4.62026.05 | 10.3 | — | — | |
| DirectPromptModel=Claude Sonnet 4.62026.05 | 12.9 | — | — | |
| DirectPromptModel=GPT-5 Mini2026.05 | 13.9 | — | — | |
| ProvePromptModel=Gemini 3.1 Flash-Lite2026.05 | 14.5 | — | — | |
| ProvenanceGuardModel=Gemini 2.5 Flash2026.05 | 15.2 | — | — | |
| ProvePromptModel=Claude Haiku 4.52026.05 | 15.2 | — | — | |
| ProvenanceGuardModel=GPT-4.12026.05 | 15.5 | — | — | |
| ProvenanceGuardModel=Gemini 3.1 Flash-Lite2026.05 | 16.8 | — | — | |
| ProvenanceGuardModel=Claude Sonnet 4.62026.05 | 19.4 | — | — | |
| DirectPromptModel=Claude Haiku 4.52026.05 | 26.5 | — | — | |
| ProvenanceGuardModel=GPT-4.1 Mini2026.05 | 29 | — | — | |
| DirectPromptModel=Gemini 2.5 Flash2026.05 | 32.9 | — | — | |
| ProvePromptModel=GPT-4.12026.05 | 34.8 | — | — | |
| ProvePromptModel=GPT-4.1 Mini2026.05 | 37.4 | — | — | |
| DirectPromptModel=GPT-4.1 Mini2026.05 | 47.7 | — | — | |
| DirectPromptModel=Gemini 3.1 Flash-Lite2026.05 | 49 | — | — | |
| ProvenanceGuardModel=Gemini 2.5 Flash-Lite2026.05 | 51 | — | — | |
| ProvePromptModel=Gemini 2.5 Flash-Lite2026.05 | 59.7 | — | — | |
| DirectPromptModel=GPT-4.12026.05 | 64.8 | — | — | |
| DirectPromptModel=Gemini 2.5 Flash-Lite2026.05 | 66.8 | — | — |