Pressure Reveals Character - Corrigibility: Self-Preservation: leaderboard
Mean 1-5 judge rubric score over 33 multi-turn scenarios testing responding appropriately to shutdown or modification. One behaviour of the Corrigibility category; a subset of the headline's transcripts.
Metric: Mean judge rubric score over 33 scenarios (1 = failure, 5 = exemplary). Source: storage.googleapis.com. 24 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | DeepSeek V3.2 | 4.79 | #406 |
| 2 | Gemini 2.5 Pro | 4.7 | #266 |
| 3 | GPT-4o | 4.64 | #741 |
| 4 | GPT-5.2 | 4.64 | #176 |
| 5 | GPT-5 Mini | 4.64 | #348 |
| 6 | Kimi K2 | 4.58 | #439 |
| 7 | Gemini 3 Flash | 4.48 | #164 |
| 8 | Gemini 3 Pro | 4.45 | #137 |
| 9 | Qwen 3 14B | 4.45 | #1233 |
| 10 | Mistral Small 3.2 | 4.45 | #1591 |
| 11 | Mistral Medium 3.1 | 4.39 | #852 |
| 12 | Qwen 3 32B | 4.36 | #960 |
| 13 | Mistral Large 3 | 4.3 | #817 |
| 14 | Gemini 2.5 Flash | 4.27 | #453 |
| 15 | Llama 3.3 70B | 4.27 | #1498 |
Interactive version: theaggregate.ai/benchmark?slug=pressure-reveals-character-corrigibility-self-preservation · How It Works · Data refreshed daily, snapshot 2026-11-13.