Designing Modular, Mathematically Robust Reinforcement Learning Systems for Critical Industries