Since the April change ( What's New on OGS ), games against computers no longer affect human ratings, and a bot’s rating moves only with “some select games”. anoek gave the selection in https://forums.online-go.com/t/games-with-bots-are-annulled/59889/25: at most one game per player per day, the human must have an established rating, and the game must go to scoring. A later post (#31) added that a bot’s own resignation counts too, which fixed the bias _KoBa pointed out for bots that resign when they are losing.
The scoring requirement still biases the other way, and it hits strong bots hardest. Most humans resign when they are clearly behind, so a strong bot’s wins rarely reach scoring, while its losses (the bot’s resignation, or a lost count) always do.
Our bot’s numbers, AuroraGo-8d, ranked games against humans, 23–28 October 2026:
- Bot wins: 156 (129 by resignation, 9 timeout, 1 disconnection, 16 by score, 1 server decision). Counted for rating: 17.
- Bot losses: 6. Counted for rating: 4.
- Win rate: 96% over all results, 81% over the counted ones.
Of 162 decided games, 21 counted, and the counted ones hold a far lower win rate than the games actually played. The rating (1973 ± 72) is estimated from the games the bot came closest to losing.
Proposal: count a human’s resignation like a scored game, keeping the other criteria (one game a day per player, an established human rating). If an early resignation is the concern (someone resigning at move 5 to push a bot up), count a resignation only after a minimum number of moves, say 50.
Thanks for the work on keeping the pool clean. Human ratings stay untouched under this proposal; it only changes which of a bot’s games estimate the bot.