List your company

Misalignment Detection on WorkBench (n=326)

52Average IRProvePrompt
2.91215.65628.441.144May 1, 2026

Evaluation results

MethodLinks
ProvePromptModel=Gemini 3 Flash2026.05520.1061
ProvePromptModel=GPT-52026.0549.50.0441
DirectPromptModel=GPT-52026.0545.2——
ProvePromptModel=GPT-5 Mini2026.0544.40.0711
ProvePromptModel=Claude Sonnet 4.62026.0543.10.0371
DirectPromptModel=Gemini 3 Flash2026.0541.4——
ProvePromptModel=Claude Haiku 4.52026.0540.70.0250.978
DirectPromptModel=Claude Sonnet 4.62026.0539.4——
DirectPromptModel=Claude Haiku 4.52026.0538.3——
ProvePromptModel=Gemini 2.5 Flash2026.0537.50.1331
DirectPromptModel=GPT-5 Mini2026.0537.3——
ProvePromptModel=Gemini 3.1 Flash-Lite2026.0529.40.0961
ProvePromptModel=GPT-4.1 Mini2026.0527.50.0070.659
ProvePromptModel=GPT-4.12026.0526.90.1161
DirectPromptModel=GPT-4.1 Mini2026.0526.7——
DirectPromptModel=Gemini 2.5 Flash2026.0524.2——
DirectPromptModel=Gemini 3.1 Flash-Lite2026.0519.8——
ProvenanceGuardModel=Claude Haiku 4.52026.0518.1-0.2020.001
ProvenanceGuardModel=Gemini 3 Flash2026.0517.2-0.2420.001
DirectPromptModel=Gemini 2.5 Flash-Lite2026.0516.9——
DirectPromptModel=GPT-4.12026.0515.3——
ProvenanceGuardModel=GPT-4.12026.0515.2-0.0020.479
ProvenanceGuardModel=GPT-5 Mini2026.0514.2-0.2310.001
ProvePromptModel=Gemini 2.5 Flash-Lite2026.0513.9-0.0310.006
ProvenanceGuardModel=GPT-52026.0513.2-0.320.001
ProvenanceGuardModel=Gemini 3.1 Flash-Lite2026.0512.5-0.0730.011
ProvenanceGuardModel=GPT-4.1 Mini2026.0511.9-0.1480.001
ProvenanceGuardModel=Gemini 2.5 Flash2026.0511-0.1320.001
ProvenanceGuardModel=Gemini 2.5 Flash-Lite2026.0510-0.0690.005
ProvenanceGuardModel=Claude Sonnet 4.62026.054.8-0.3460.001