CausalNav: Certified Causal World Models for Robust Control.

For physical AI systems, a world model is only valuable if it influences agent actions and safe if it knows when to abstain. CausalNav addresses this dual requirement by integrating a controller with a signed, action-conditioned transition graph over identified state coordinates. During deployment, CausalNav simulates intervention sequences, translating objective errors into policy-logit advice. This advice is only adopted if a scale-free predictive-reliability certificate, a policy-margin gate, and an argmax-agreement gate all pass; otherwise, the system reverts to its model-based base controller. Evaluated against nine baselines on CartPole-v1 and discretized Pendulum-v1 under physical-parameter shifts, CausalNav achieved the best average rank. A significant diagnostic finding was that while the learned graph recovered structure well, per-seed structural fidelity did not correlate with control benefit. Crucially, the certificate-based abstention mechanism proved vital for safety, particularly in Pendulum tasks where forcing the planner on incurred performance costs. This research suggests that certified abstention, rather than merely improved prediction, is paramount for safely deploying world models in physical AI.