Artifacts
What each model built on each task, shown on the held-out cases it was graded on.
Stateful Parquet table
Every model optimized an incrementally updated Parquet table whose contents and query answers must match the reference, so what differs is latency, memory and bytes written. The race replays the graded latency on the 3 hidden test streams the reward is computed on; memory and writes are on the results board.