When RL training only verifies the final outcome, there is no incentive for a model to keep its intermediate reasoning human-readable