Conformal Risk-Calibrated Latent World Models with Runtime Assurance Under Distribution Shift

Abstract

Learning-enabled autonomous controllers may fail when a compact world model remains confident outside its training distribution. We present the conformal disagreement-regulated latent world model (CDR-LWM). The method converts ensemble disagreement into a risk-calibrated, state-dependent uncertainty radius for robust-envelope model predictive control. A split-conformal score scales next-state residuals by ensemble dispersion. A runtime governor then replaces the planned action with a stabilizing backup when the predicted envelope approaches a safety boundary or the ensemble novelty exceeds its empirical nominal quantile. We evaluated CartPole and torque-limited Pendulum stabilization using 15,000 locally generated transitions per task. The protocol used episode-disjoint fitting, calibration, and validation partitions, together with eight paired held-out seeds under nominal, dynamics, actuator, and observation conditions. Under Pendulum actuator shift, CDR-LWM reduced unsafe steps from 7.3% to 0% and episode failures from 50.0% to 0% while intervening on 9.3% of steps. Under dynamics shift, unsafe steps decreased from 14.9% to 7.4%, and episode failures decreased from 62.5% to 12.5%, with a 30.0% intervention rate. Under CartPole observation noise, CDR-LWM completed every evaluation episode without an unsafe state, whereas nominal MPC had a 37.5% episode-failure rate. Nominal one-step joint coverage was 0.909 for CartPole and 0.882 for Pendulum. Trigger-isolating ablations showed that novelty-triggered switching produced most of the closed-loop improvement. These results provide a compact and auditable link between calibrated model disagreement and selective runtime assurance under the evaluated shifts.