Scalpel: An LLM-Team Architecture for Symbolic Regression of Physical Laws from Noisy Data

We present Scalpel, a symbolic regression system where a team of 9 LLMs from distinct architectural families discovers physical laws from noisy data without domain knowledge. The key principle — anti-incest — ensures no two models share the same architectural family, providing genuine independence of verification. On the Feynman Benchmark with 10% Gaussian noise, Scalpel recovers 14/20 laws (70%) in a single clean run on CPU with 32 GB RAM, no GPU. Exact structural matches: 4/20. Structural approximations with learned physical constants: 4/20. The system learns across runs: STOP-patterns, Matryoshka advice, and meta-patterns accumulate between launches. Fully explainable — every hypothesis, debate, vote, and formula is logged. Built by one person in 9 days with no scientific background. Code: https://github.com/Yapan777/scalpel