Physics-Informed Planning

When the Hamiltonian family is known, TAS-AI reached the target RMS threshold faster than the competing methods in the representative refinement run. Physics-informed acquisition ranks candidate points by expected information gain per unit wall-clock time. AIC-derived weights are used as a real-time proxy for model evidence during Hamiltonian discrimination. The local posterior is approximated as Gaussian, and expected information gain is computed with a Laplace-style expression for fast in-loop ranking. MCMC is reserved for batch boundaries or offline validation when local diagnostics fail. Physics-informed planning becomes useful after a plausible signal region and model family are available.