Artifacts
What each model built on each task, shown on the held-out cases it was graded on.
Chess puzzles
Qwen3-8B re-plays chess puzzles with each model's trained adapter, and with none: the moves it played and the reasoning it wrote before each one. Hidden test puzzles are 24 of the 1,000 the reward is computed on (each model's best and median attempt); public dev puzzles (best attempt) are ones agents could train on, so they are not graded.