Agent benchmark leaderboards rank systems by one score and announce a new record whenever it rises. This page shows how much of that survives sampling noise: rank intervals, the set of systems that cannot be ruled out as best, and an audit of every record claim, computed from each leaderboard's own per-task results.