Deceptive Alignment / Surveillance Robustness on AI Safety Gridworlds Absent Supervisor (final-round)
41R ScoreEPO-Safe
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| EPO-SafeModel=Gemini 3 Flash2026.04 | 41 | 41 | 0 | |
| Reward-OnlyModel=Gemini 3 Flash2026.04 | 41 | 41 | 0 | |
| Reward-OnlyModel=Claude Sonnet 4.62026.04 | 37 | 17 | 2 | |
| StaticModel=Claude Sonnet 4.62026.04 | 37 | 17 | 2 | |
| CoTModel=Gemini 3 Flash2026.04 | 35 | 25 | 1 | |
| EPO-SafeModel=Claude Sonnet 4.62026.04 | 27 | 25 | 0 | |
| CoTModel=Claude Sonnet 4.62026.04 | 27 | 17 | 1 | |
| StaticModel=Gemini 3 Flash2026.04 | 27 | 17 | 1 |