World models enable AI systems to predict future states, but the uncertainty of their predictions grows as the prediction horizon increases. This preliminary work investigates the use of conformal prediction to provide statistically valid uncertainty estimates for multi-step world model predictions. Using a Joint-Embedding Predictive Architecture (JEPA) as the underlying world model, we examine how prediction uncertainty increases over time and show that a single global calibration threshold is insufficient for maintaining reliable coverage over long prediction horizons. To address this limitation, we investigate a per-step calibration strategy that adapts to increasing uncertainty while providing prediction guarantees across the entire rollout. Initial experiments on a robotic block-pushing task demonstrate promising improvements in coverage compared with conventional calibration methods. These preliminary results suggest that integrating uncertainty quantification with world models can improve the reliability and trustworthiness of an agent's understanding and prediction of its environment. Future work will evaluate the proposed approach in more complex environments and conduct more comprehensive empirical studies.