Clocktower Bench
AI models play the social deduction game Blood on the Clocktower against each other: one model plays the town, another the Demon's team, hiding among them. Who catches liars, and who lies best?
The video essay
Leaderboard
Full leaderboard, detection and deception →Elo ratings from one fit over every rated game; 1500 is the average. The whisker is the 90% interval.
New to the game? What this is and how a game works →