When a world model detects distribution shift, what should the agent do? Current approaches offer a binary choice — continue or retrain — ignoring the spectrum of adaptive responses available. We present a unified framework combining the Homeostatic Hamiltonian Agent (HHA) with the Model Invalidity Test (MIT). HHA introduces a stress-energy functional that governs adaptation intensity through a dynamic gain parameter gamma, enabling three distinct response modes: freeze (preserve the current model), adapt (incremental update), or rebuild (full reconstruction). MIT provides the conjunctive decision rule that triggers these responses, testing both prediction accuracy and physics-model alignment. Across six benchmarks covering panic-freeze scenarios, noise specificity, structural recall, response taxonomy, ablation studies, and SOTA comparison, the HHA+MIT system demonstrates appropriate response selection with measurable separation between response modes. The framework provides a principled alternative to threshold-based switching by grounding adaptation decisions in physical quantities. This archive contains the full LaTeX source, compiled PDF (20 pages), and all benchmark scripts with reproducibility instructions.
Régis RIGAUD (Sat,) studied this question.