Artifacts
What each model built on each task, shown on the held-out cases it was graded on.
Draft heads
Each model trained a small draft head that guesses SmolLM3-3B's next seven tokens; the target keeps the guesses it agrees with, plus one token of its own. Hidden test prompts are the private ones the reward is computed on (six shown, for each model's best and median head); public dev prompts are the 20 the agents could test on, so they are not graded.