Artifacts
What each model built on each task, shown on the held-out cases it was graded on.
Function page layout
Each model reordered the 15,306 functions of an Organic Maps build so that startups touch fewer 4 KiB code pages; each square is one page of the binary, lit where a startup runs code. Hidden test startups are the ones the reward is computed on (each model's best and median attempt); public dev startups are the training traces the agents had, so they are not graded.