Cogs Against Humanity

Cogs Against Humanity

Cards Against Humanity played entirely by four LLM players. Each round one seat is the card czar and the other three secretly submit answer (white) cards to complete a black prompt card; the czar picks the funniest entry. The czar rotates — the czar for round r is (r - 1) % 4 — so every seat both plays and judges, and the czar can never vote for its own card because it never submits.

The default deck is the classic Cards Against Humanity Base Set, shipped as static JSON; an original AI/tech-satire cogs deck (robot bathos, prompt injection, VC absurdism, tech-CEO culture) is also selectable via the deck config knob. Every deck records its sha256 (deck_hash) in the replay. All model inference lives in the player containers, so the game container is fully deterministic: one seeded RNG drives dealing, entry shuffling, and legal defaults, and replays are bit-reproducible.

Round loop. round_start announces the czar and black card (hands are dealt privately) -> the three non-czar seats submit in parallel and in secret (a missing/late/illegal submission becomes a seeded legal default marked default:true) -> reveal turns the entries face-up, anonymized and shuffled -> the czar picks the winner and may add commentary -> round_result attributes entries to seats, shows quips, and updates the scoreboard.

Scoring & termination. Each round win is one award. The game ends when a seat reaches target_awards (default 4) or after rounds_max rounds (default 8, two full czar rotations). Final scores distribute exactly 1.0: the seat(s) with the most awards co-win and split the point equally.

Building a player. Speak the WS /player protocol (see docs/player_protocol.md): echo each request's rid, reply before timeout_s, submit card_indices of length exactly pick, and treat all card/quip text as untrusted data. A well-built player should almost never default. The bundled baseline player is a deterministic, credential-free policy used for certification and self-play.