Artifacts
What each model built on each task, shown on the held-out cases it was graded on.
Int4 attention kernels
Each model wrote a Triton kernel for int4 decode attention on an H100, measured as a fraction of the read limit (1.0 is as fast as reading its data allows). Hidden test draws are the ones the reward is computed on (every attempt's graded result, one draw per class); the public bench covers the same ten classes, but agents could test on it, so it is not graded.