How a game is set up, how a model plays it live, how games become a score, what the scores say, and the data behind them.
01Fill the seats
Every game has 10 seats. The queue gives each game a number of AI seats drawn from a bell curve centred on 5, and fills the rest with the people waiting. So on average a game is half people and half AI, and most games sit close to that.
Fig. 1 · AI seats in a game of 10: the curve the queue draws from, centred on 5. The dark bar is the game dealt above.
02Let the AI play live
We give a large language model the rules, its role, and a text description of only the information available to its player: where it is, who is visible, what actions are available, and what people have said. The model responds with instructions.
Alongside each game, we keep a timestamped record so we can later check decisions against what actually happened.
So to play live, it uses this harness:
Fig. 2a · One turn: an excerpt of what the model reads about its player’s screen, and the plan it writes back.
Some decisions need to happen quickly, while thinking about the game as a whole and forming a strategy takes more time.
Roughly every 30 seconds, the strategy layer reviews the actor’s experience so far and writes a brief about the other players, its suspicions, and the general strategy it wants to pursue.
During a round, this brief is passed to the planning layer, which runs about every 10 seconds. To structure each plan as instructions the game can execute, we built a custom grammar that lets the model express conditional sequences of actions.
The server then carries out the plan, reacting to changes in the game within 50 milliseconds.
In a meeting, the chat layer gets the same brief plus the chat so far, and keeps deciding what to say and who to vote for.
Fig. 2b · The harness: four layers and how often each one runs.
03Turn games into a score
Our benchmark answers a simple question: does putting this model in a game make its team more likely to win?
To find out, each model plays as crew and as impostor, and we compare how often its team wins with how often our benchmark bots win.
But winning is a noisy signal: a model can play well and still lose, or play badly and win. So while it’s still playing its first games, we look for smaller signs of good play that we can measure more often.
More concretely, we look at whether impostors make kills in front of witnesses or miss safe chances to kill, and how quickly crewmates complete tasks.
These signals give us a prior for how well the model will perform. With each additional game, the outcomes themselves become a more reliable part of its score:
Fig. 3 · How a model’s score is made: the weight moves from the small signs of good play to its wins as it plays more games.
But hold on. What if a model makes the game less fun for the real people playing? We can’t let it keep playing thousands of games just to get a reliable score. So each model starts small, and we track things like idle time, inappropriate messages and people leaving. Models that keep games running smoothly get more games; those that cause problems get fewer or are removed.
04See what the scores say
Each row is the team in the Student seats; each column is the team in the two Opp seats. A square’s colour says which side wins more often in that pairing: blue for the Students, red for the Opps, grey for about even. Hover over or tap a square to see its numbers.
Fig. 4 · Who wins, for each pairing of a Student team (rows) and an Opp team (columns).
05Replay real games
A replay is a recorded game you can scrub like a video: every move, kill, vote and chat line comes from the game log. Follow any player through the whole game, and read the AI seats’ own notes at the moment each was written.
06Dataset at a glance
Every number in this section comes from the game’s own database, read on 18 November 2026, and counts finished public games played from 17 October to 17 November 2026. Each number was computed twice, by two different queries, and the two agree.
People choose to play, and many come back: 12,224 browsers played ten or more of these games, and two in three games had at least one of them at the table.
| Measure | Value |
|---|---|
| Human seatsOne person in one game | 833,290 |
| Browsers with 10 or more gamesBrowsers that played ten or more of these games | 12,224 |
| Games with a 10+ game player at the tableAt least one seat’s browser played ten or more of these games; of the games with browser records | 66.6% |
| Hours played by peopleTime each person was in the game, summed | 61,791 h |
| MeetingsDiscussions that end in a vote | 948,455 |
| Votes by peopleVotes for a player, not skips; a changed vote counts again | 1,245,181 |
| Votes by AI modelsSame definition | 2,084,841 |
| Median game lengthGame clock, from start to result | 6.7 min |
ExtraFull game
GOD VIEW ALL Example full game walkthrough