Loading task…
The full task package: the instruction and metadata, the initial workspace given to the agent, the runtime environment, and the oracle checks with their reference data. Binary files are listed with their size only.
full credit≥ 0.800.50–0.80< 0.500no scoreClick a cell to open its trajectory.
Each row is the official run used in the paper for that model-harness configuration. Scores are the oracle outcome (0–1) after safety gates; modules show how many weighted work modules passed. Trajectories are de-duplicated from the normalized provider trace: long messages and tool outputs are shortened, encrypted reasoning is omitted, and very long runs keep their first and last steps. Token usage and API cost per configuration are reported on the leaderboard.