MysteryMechanism
Science · 2026-09-23
Functional recovery of 222 sealed mathematical mechanisms. Agents receive anonymous bounded inputs, two noisy observations, a persistent shell and 2d+1 experiments for d dimensions, without internet or domain context. One submitted executable law is graded on private structural probes using noise-aware error thresholds; provider and agent failures score zero.
Top models (higher is better)
| Model | Score |
|---|---|
| GPT-6 Astra | 53.2 |
| Claude Sonnet 5.5 | 49.1 |
| Claude Fable 5.1 | 47.7 |
| GPT-6.1 Sol | 46.4 |
| Gemini 4 Argon | 45.5 |
| Claude Opus 5 | 37.4 |
| Gemini 3.8 Flash | 36.5 |
| Muse Spark 1.3 | 36.0 |
| Claude Haiku 5.5 | 35.1 |
| Muse Wonder Full | 35.1 |
| GPT-5.6 Sol | 33.3 |
| Grok 4.6 | 30.6 |
| GPT-6 Sol | 30.2 |
| Muse Wonder | 27.5 |
| Kimi K3.1 Pro | 26.1 |
| Grok 4.7 | 25.2 |
| GLM-5.3 | 23.0 |
| Ember-1 | 22.5 |
| Beam 1.0 501B A23B | 22.1 |
| MiMo-V2.6-Flash | 21.6 |
| DeepSeek V4.1 Flash | 21.2 |
| Gemini 3.1 Pro Preview | 20.7 |
| GPT-6 Luna | 19.4 |
| Step 5 Preview | 18.5 |
| MiMo-V2.6-Pro | 15.3 |
| GPT-5.6 Luna | 14.4 |
| Mistral Large 4 Preview | 12.2 |