A live multiplayer Werewolf game built as a research platform for studying how trust, suspicion, and deception unfold between players — and, eventually, between LLM agents.
A fully playable, browser-based version of the social deduction game Werewolf, designed from the ground up as a research instrument. Players join a room with a code, get a secret role, and live through alternating night and day phases of accusations, votes, and quiet eliminations. The roles in play: • Werewolves — secretly pick a nightly victim and must blend in by day • Seer — investigates one player per night, learning only "werewolf" or "not a werewolf" • Witch — holds two one-shot potions, one to save the night's victim and one to poison any player • Villagers — no powers, just their voice and their vote • Mayor — elected on the first morning, vote counts double for the rest of the game Beyond the game itself, there is a research admin layer that captures the full social trace of each match: who said what, who voted for whom, who suspected whom and when. The goal is to use this as a substrate for studying trust dynamics — first between human players, then between LLM agents dropped into the same game with the same rules.
The hardest part is that Werewolf only works if the social fabric of the game does. That meant getting the realtime layer rock-solid (no missed messages, no desynced phases, no awkward reconnections mid-vote), designing role flows that are private enough to preserve hidden information but legible enough that new players don't get lost, and structuring the game state so every action is cleanly recorded for later analysis without leaking information back into the game. On the research side, the open question is how to turn thousands of chat messages and votes into something quantitative — a measurable signal of trust that holds up across both human and LLM players.