Robotics RL lacks a universal post-training recipe like LLMs have. Current robotics models show complex behaviors but lack reliability for real-world deployment. LLMs solved this with supervised tuning, RLHF, and verifiable rewards. Robotics needs a similar recipe but with higher reliability standards. Existing RL methods for robotics struggle with expensive samples and long-horizon credit assignment. Frontier robotics models use diffusion for action generation, which older RL methods cannot handle. Value-based RL faces stability issues at scale with large models. A universal recipe requires both a stable algorithm and standard training protocols. The gap between craft and recipe must be closed for deployable robotics.