HALT: Certifying Trust Horizons in Action-Conditioned Robotic World Models

Model-based robot controllers treat the imagination horizon as a fixed hyperparameter, yet the number of steps for which a world model remains dynamically valid is state-, action-, and context-dependent. This paper formalizes this gap as silent dynamical failure and proposes HALT (Horizon-Adaptive Latent Trust), a wrapper that converts any latent or video world model into one that reports a certified, state-dependent trust horizon with a distribution-free guarantee. HALT couples a monotone quantile head over deployment-available features with simultaneous split-conformal calibration. A truncated-rollout simulation lemma links the certificate to control performance, and an adaptive-conformal variant handles dynamics shift with delayed feedback. The method is accompanied by six action-conditioned metrics and PHANTASM, a six-axis stress suite. Across five environments, ten seeds, and three world model classes, HALT maintains coverage within one percent of the target while providing usable certified horizons in contact-rich tasks, outperforming fixed-horizon baselines by fifteen to thirty percent in success rate and reducing unsafe contacts by forty to sixty percent. While prior work has explored uncertainty estimation for world models, none has provided a distribution-free certificate for the imagination horizon in task-relevant units. HALT offers the first such guarantee for action-conditioned world models.