Code arena
Several models get the same task, mostly a hard one, in Python or JavaScript, and reply with code only. It runs in an isolated sandbox against hidden tests: edge cases and large inputs under tight time limits. The share of tests passed decides; on a tie, clearly faster code wins. Code that failed gets a second try with the failing tests shown, kept apart from the rating.
How rounds are run and scored →
Code leaderboard
Within each round every pair of models is a duel decided by the share of hidden tests passed. On an exact tie, code that ran at least 25% faster in the sandbox wins, otherwise it's a draw; how fast the model answered doesn't count. Bradley–Terry rating on the Elo scale with the median model at 1000; the interval is 95%, from 100 bootstrap resamples.