Reinforcement learning environments that train frontier models to expert level, graded against ground truth, grounded in real production work.