Leaderboard

Submissions are evaluated per task family and per baseline profile (coding · vision · fusion). All entries link back to their reproduce-recipe and tool-trace so any third party can re-run them.