Scalpel: 9-LLM team guiding PySR to find physical laws — 3/5 Feynman laws on noisy data

[Yapan777]

Hi! I built Scalpel — a system where a team of 9 LLMs from different architectural families guides PySR to find physical laws from noisy data without domain knowledge.

Key idea: anti-incest. Each model comes from a different family — DeepSeek, Qwen, Phi, Mistral, Granite, LLaMA, Gemma, Yi, InternLM. They debate before PySR runs, vote independently after (Matryoshka audit), and remember past failures across runs via STOP-patterns.

Results on Feynman Benchmark, 10% Gaussian noise:

• 3/5 laws recovered
• CPU only, 32 GB RAM, no GPU
• Full reasoning log for every decision
• System learns across runs

Built in 9 days from zero knowledge of symbolic regression.

GitHub: https://github.com/Yapan777/scalpel
Preprint: https://doi.org/10.5281/zenodo.19476680

Would love feedback on improving the PySR integration!