CauCity-WM: an intervention-consistent generative world model for counterfactual urban dynamics and risk-aware decision making

Urban decision making requires more than accurate forecasting: it requires estimating how the same realized episode would evolve under an action that was not taken. We introduce CauCity-WM, an intervention-consistent generative world model whose methodological core is a noise-preserving twin-world transition. The model abducts an episode-specific structural disturbance from the factual trajectory and then reuses both that disturbance and the same diffusion base-noise path in factual and counterfactual rollouts, changing only the intervention. This defines a unit-level counterfactual semantics for a graph-conditioned diffusion world model rather than treating causality as an auxiliary prediction head. Mechanism-gradient alignment promotes transfer across cities and regimes, sparse intervention Jacobians constrain action propagation, and risk-sensitive planning optimizes CVaR-regularized actions through the same transition law. Across seven real urban forecasting benchmarks, CauCity-WMreduces average MAE relative to the strongest forecasting baseline. In counterfactual evaluation, it achieves an E-MAE of 2.61 and PEHE of 3.29, compared with 3.22 and 4.07 for the strongest adapted causal baseline, while attaining 84.6% effect-sign accuracy and 89.6% coverage of nominal 90% effect intervals. In calibrated SUMO scenarios, it reduces mean travel time to 27.1 min and constraint violations to 1.9%. These results support intervention-consistent urban world modeling as a practical basis for counterfactual analysis and risk-aware decision making. Code is available at https://anonymous.4open.science/r/CauCity_WM-C563 .