Minimum Description Length Separates Scientific Laws from Statistical Noise

Using Minimum Description Length (MDL) as an unsupervised epistemic grader, a self-improving science agent successfully distinguished canonical physical laws from statistical noise. The method recovered known laws like the Balmer/Rydberg series with high significance while revealing that 86% of the agent's discovered relations provided no compression gain.