Share of rubric items satisfied (%) on the 220 public questions, for the base model and the chosen checkpoint of each stage. Dashed lines: The two Samaya System configurations on the public leaderboard.
Figure 8: Each stage starts from the previous stage's checkpoint. The recipe adds 16.4 pp over the base model, and the two RL stages improve the SFT checkpoint by 4.1 pp (54.8% → 58.9%); both RL checkpoints score above the two Samaya System configurations, and the online GRPO checkpoint improves on the previous high-effort configuration in production by 2.9 pp (58.9% against 56.0%).