Five SFT runs that scored zero, a failed RL attempt, switching models, and learning to decompose the problem.

Procedural data generation, testing existing models, and discovering the connection to robotics.