Traverse.so builds reinforcement learning environments and training data that help frontier AI labs teach models taste, judgment, and long-horizon work.