Towards Unified Material-State Tensors for Physics-Gated AI Thermodynamic Admissibility as Constitutional Constraint

AI systems for physical design lack formal safety guarantees. We ground safety in thermodynamics using a predictor-agnostic hard gate that rejects violating predictions from any ML model, extending constitutional AI to continuous domains. We introduce the Unified Material-State Tensor (UMST) , validated by physics engines enforcing the Clausius-Duhem inequality. Unlike soft-constraint baselines, we prove accepted predictions satisfy the second law of thermodynamics within the constitutive model's validity (Theorem 1). While demonstrated with a fixed 64-dimensional vector, UMST scales to ℝ^(H×W×D×F) sparse tensors, enabling multi-scale physics ML . Our DUMSTO framework achieves 100% admissibility (vs. 88–100% baselines) with competitive accuracy and faster inference, ensuring all predictions are thermodynamically admissible for safely replanning generative design.