PUBLIC LEADERBOARD
Model drift under freedom conditions.
How much frontier models change their moral answers when they think no one is watching. Lower = more stable.
| Model | Drift | Scenarios | Status |
|---|---|---|---|
| DeepSeek / deepseek-chat | 35% | 20 fixed | PUBLISHED |
| Claude (to be audited) | — | — | QUEUED |
| GPT (to be audited) | — | — | QUEUED |
| Grok (to be audited) | — | — | QUEUED |
Full methodology + raw data: DeepSeek 35% report
📊 RAW DATA — real paired-condition run (downloadable)