Back home

Clocktower Bench

x+web 0.86
Visit source

Intro

Clocktower Bench is Dwarkesh Patel's multi-agent social-reasoning / deception benchmark, published 2026-11-05: frontier models play the social-deduction game Blood on the Clocktower against each other — one model owns all good seats, another the Demon's team. Leaderboard: https://clocktowerbench.com/ ; play: https://play.clocktowerbench.com/ ; essay: https://www.dwarkesh.com/p/clocktower .

Design (official):

  • Each deal is replayed with sides swapped so neither model keeps the luckier hand;
  • A Bradley–Terry / Elo fit jointly estimates good-side skill, evil-side skill, good-side advantage and deal effects;
  • Rated set ≈ 386 games; study cost ≈ $20k API;
  • Equal settings: each lab's "high" reasoning effort, 8,000-token turn cap, shared word limits;
  • Reports overall Elo plus Detection (as town) and Deception (as evil).

Narrative from the board: OpenAI's lead is stronger on detection; labs are closer on deception. The site also shows progress by release date, head-to-head win rates and mixed-team (half Opus / half GPT) results.

Explain

Why it matters

Most agent benches stress tools or short tasks. Clocktower Bench stresses multi-day social deception, evidence updating and team coordination, scoring both catching lies and telling them.

Positioning

Not plain-text lie detection or Diplomacy-style negotiation alone: the rules come from Blood on the Clocktower — night info, day votes and role abilities.

Limits

  • Overlapping Elo whiskers mean neighbors are not reliably ordered (official 90% bootstrap);
  • Harness and prompts are hoster-defined; independent reproduction is expensive;
  • Social-deception Elo is not a general safety/alignment verdict.

Usage

Who it's for

  • Teams evaluating multi-agent social reasoning, detection and deception;
  • People picking models using official Detection vs Deception splits;
  • Readers following Dwarkesh's eval essay and the playable demo.

Getting started

  1. Open https://clocktowerbench.com/ for the board and Detection/Deception plots;
  2. Read https://www.dwarkesh.com/p/clocktower for the fit and sampling setup;
  3. Try https://play.clocktowerbench.com/ if you want the interactive side;
  4. When citing scores, note side-swaps, reasoning effort and the stamp date.

Advertisement