All results
Every run of the study.
How the verified count breaks down
Ranking
Only time was held equal; reasoning effort and harness differ per agent (Agents).
Best run per agent and task
Best run: an agent's verified final with the fewest nodes on a task.
Scroll sideways for every column →
▲ the smallest best run on the task.
Every best run and every task
Agents
Every run, by task
Each mark is a run. The dashed line shows the automatic sharing pass, which writes each repeated subterm of the original once.
Automatic sharing pass on the original proofs
How the runs ended
By task, agent and time limit
Final node counts of the verified runs: the best run (fewest nodes) and the median. “No score” counts the runs without a verified score.