When Physics Disagrees: The Resonance Signal as a Diagnostic for Model-Physics Alignment

Detecting when a world model diverges from physical reality is critical for safe model-based reinforcement learning. Current approaches treat prediction error as a monolithic signal, conflating noise, drift, and structural failure. We introduce Resonance-Based Detection (RBD), a diagnostic framework that compares model predictions against a physics prior to produce a resonance signal R ∈ 0,1 measuring model-physics alignment. The blending mechanism between physics prior and learned model enables graceful degradation: when the model fails, the system falls back to physics-based predictions rather than failing catastrophically. Experiments on CartPole and spring-mass systems demonstrate that RBD distinguishes structural failures from noise with high specificity, achieving detection within 5 timesteps while maintaining zero false positives on noise-only controls. The vectorial extension localizes failures per dimension, enabling targeted rather than full model reconstruction. This archive contains the full LaTeX source, compiled PDF (17 pages), and reproducibility scripts including all experiments and baselines.