Bounding-Box Inference for Error-Aware Model-Based Reinforcement Learning