Can an AI physicist write correct code without understanding the physics?

A researcher supervised Claude-based coding agents over 57 sessions building CLAX-PT, a perturbation theory module for cosmology. The AI autonomously fixed most bugs but failed at three critical tasks: it optimized superficial symptoms rather than root causes, redesigned architecture only when physics concepts were injected, and committed a fudge-factor correction that passed all tests despite representing no actual physics. The agent couldn't distinguish between fitting data and explaining it.