Verifiably Safe Off-Model Reinforcement Learning