CRED v0.92
Science · 2026-07-20
CRED v0.92 evaluates automated verifiers on locating known errors injected into 200 copies of 100 self-reproducing empirical policy papers. Each configuration reviews the manuscript, rendered tables, analysis code, bibliography, and reproduction diagnosis; deterministic string matching credits findings that quote the changed content. The headline score is pooled single-error detection recall, with failed calls and refusals counted as misses.
Top models (higher is better)
| Model | Score |
|---|---|
| GPT-6 Astra | 99.5 |
| GPT-5.6 Sol | 99.0 |
| GPT-5.6 Luna | 97.5 |
| GPT-5.6 Terra | 97.0 |
| GPT-5.5 | 96.5 |
| Claude Fable 5 | 96.5 |
| GPT-5.6 Luna Pro | 91.0 |
| GLM-5.3 Flash | 90.0 |
| GLM-5.2 | 89.5 |
| GPT-6 Luna | 87.5 |
| Opus 4.8 | 85.0 |
| Sonnet 4.6 | 83.5 |
| DeepSeek-V4-Pro | 83.5 |
| DeepSeek-V4-Flash | 80.5 |
| Gemini 3.8 Flash | 79.5 |
| Sonnet 5 | 77.5 |
| Gemini 3.7 Flash | 74.0 |
| Gemini 3.5 Flash | 68.5 |
| Hy3 preview | 50.5 |
| Gemini 3 Flash Preview | 47.0 |
| Mistral Large 3 675B Instruct 2512 | 32.0 |
| Mistral Small 3.2 24B Instruct 2506 | 19.0 |
| GPT-4.1 | 14.0 |
| GPT-4o (source unspecified) | 11.5 |