Leaderboard
Ratings on an Elo-like scale (+400 points = 10 to 1 odds), anchored on bots of fixed strength: the random player is 0. One table per board size, one row per model and aid level, with the 95% interval. A wide interval means few games: read it before the rating.
Does the help help? The aid effect
For each model, its rating with an aid minus its rating bare (L0), on the same board size. Above zero, the aid helped.
Where the ratings stand
The leaderboard is computed from tournament games only. 0 finished games so far: 25 adjourned (stopped at a move limit, no winner); 84 played before the tournament format, without the registry ids that tie a player to a model; 5 flagged invalid (result caused by a referee bug since fixed); 2 without a scenario file. None of them can be rated yet, so the table above is empty on purpose rather than filled with one-game guesses.
Next: the anchor bots play each other to fix the scale (random, greedy, Monte Carlo, KataGo at one visit, KataGo imitating 20 kyu and 10 kyu players), then every model plays them bare, then each model plays itself with each aid. See the method.
Meanwhile, the reference for language models on 9×9 Go is GoBench (Roland Gao); its results are listed per model on the models page. We plan to share its KataGo reference levels as anchors so the two scales can be compared.
Method: Bradley-Terry, anchored on bots; White-advantage term; one fit per board size. Intervals from 1000 bootstrap resamples. Updated 2026-11-14T15:29:08+00:00.