PhysDox (Self-Consistency): leaderboard
Metric: Stage-1 severity-detection macro-F1 (0-100) on the 683-protocol expert-curated Gold set of biomedical physiological-sensing protocols (six domains): classify each protocol as valid, minor issue or fatal physical-feasibility violation; self-consistency: majority vote over three temperature-0.7 samples; higher is better. Source: arxiv.org. Saturation forecast: Around 2031. 6 models tracked.
Top models
| # | Model | Score |
|---|---|---|
| 1 | DeepSeek V4 Flash | 53 |
| 2 | Claude Opus 4.7 | 51.4 |
| 3 | MiMo-V2.5 | 46.3 |
| 4 | DeepSeek V4 Pro | 45.5 |
| 5 | MiMo-V2.5-Pro | 44.7 |
| 6 | GPT-5.5 | 38.5 |
Interactive version: theaggregate.ai/benchmark?slug=physdox-self-consistency · How It Works · Data refreshed daily, snapshot 2026-09-29.