STORM-World: A Physics-Informed, Intervention-Aware World Model for Counterfactual Storm Simulation

World models for physical AI should represent not only observations, but also the latent physical state of an environment and how that state changes under interventions. We introduce STORM-World, a physics-informed, intervention-aware latent world model for coupled tropical-cyclone, infrastructure, population-exposure, and response dynamics. STORM-World combines a physics-regularized graph–operator encoder with recurrent latent state-space dynamics and branch-state interventions that propagate through learned open-loop rollouts rather than directly editing model outputs. We evaluate the model using a five-axis protocol covering predictive, physical, probabilistic, intervention, and decision fidelity. Physics regularization reduces all six atmospheric residuals, including a 4.5$\times$ reduction in diffusion residual, while increasing track error by 14%, showing that forecast accuracy and physical consistency can move in opposite directions. Evacuation-timing interventions produce a monotone dose–response across all 10,000 bootstrap resamples, while action rankings are preserved with mean Spearman $\rho=0.975$, Kendall $\tau=0.958$, and 235/240 correct pairwise comparisons. However, four infrastructure and resource intervention channels are null or wrong-signed, and stochastic rollout intervals become increasingly over-confident with horizon. These failures are reported as first-class results. Our findings suggest that physical world models, embodied or otherwise, should be evaluated for physical consistency, calibrated uncertainty, controllability, and decision fidelity—not forecast accuracy alone. Here, counterfactual refers to model-generated intervention scenarios rather than identified real-world causal effects.