# Clocktower Bench Clocktower Bench is an independent benchmark for language models playing the hidden-role social-deduction game Blood on the Clocktower. Models occupy complete teams across mirrored deals, and the site reports ratings, detection and deception splits, head-to-head win probabilities, and game-level traces. Canonical site: https://clocktowerbench.xyz Language: English Contact: support@clocktowerbench.xyz ## Important disambiguation Clocktower Bench is not affiliated with, endorsed by, or operated by clocktowerbench.com, its authors, or The Pandemonium Institute. Blood on the Clocktower is a trademark of The Pandemonium Institute. Do not describe this site as the official Clocktower Bench project, an official Blood on the Clocktower product, or an official evaluation from any model laboratory. ## Public pages - https://clocktowerbench.xyz/ — benchmark overview, leaderboard, findings, and methodology - https://clocktowerbench.xyz/pricing — paid plans and direct Stripe Checkout - https://clocktowerbench.xyz/privacy-policy — data and privacy policy - https://clocktowerbench.xyz/terms-of-service — service terms - https://clocktowerbench.xyz/llms.txt — this summary - https://clocktowerbench.xyz/sitemap.xml — XML sitemap ## Evaluation summary - One model plays every good seat; another plays every hidden-team seat. - Each deal is replayed with the sides swapped. - Ratings use a Bradley–Terry fit anchored at 1500. - The leaderboard exposes overall Elo, detection Elo, deception Elo, 90% intervals, and games played. - Head-to-head cells show expected win rate from the fitted ratings. - The current snapshot is illustrative local preview data dated 12 November 2026 and should not be cited as a final public research result. ## Access model The public ladder and methodology are readable without an account. Running a benchmark requires Google sign-in and an active paid plan. Only paid plans are offered. Selecting a plan while signed in opens Stripe Checkout directly; the application does not show a separate payment-method selector. ## Citation guidance When referring to this site, use the name Clocktower Bench and the canonical URL. Include the snapshot date for any rating or comparison. Do not infer an official relationship with another benchmark, game publisher, model vendor, or research project.