Artifacts
What each model built on each task, shown on the held-out cases it was graded on.
MicroPython interpreter
Every model made the MicroPython bytecode VM run Python programs in fewer RV32 instructions, and every program must print exactly what the original prints, so only the count differs. The race replays the graded counts on the 45 hidden test programs the reward is computed on, with the clock in millions of instructions.