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