One word.
One outsider.

In elimination mode, vote after each round. A civilian leaving starts another round; a tied vote removes nobody. Catch the undercover to win as a civilian. The undercover wins when one civilian remains.

In three-round mode, the final vote must catch the undercover. A wrong target or tied vote gives the undercover the win. You may accuse someone early: a correct accusation wins for the civilians; a wrong accusation eliminates you and play continues unless a victory condition is reached. AI players may accuse when target support reaches 80%, but can keep observing. Support is not measured accuracy.

Each AI receives its own word, all earlier public clues and events, and its own action history. Jev compares bounded role hypotheses before selecting a prepared clue, a legal vote or an eligible accusation. It can try to blend in when it suspects itself. Code handles dealing, turn order and victory. Repeating any earlier clue, including a paraphrase, can make a player suspicious; an actual new detail matters.

When the game ends, see every player's word and role. Replay the visible evidence, options and returned scores for each decision, or download the recorded log. Scores are not calibrated identity probabilities, and the replay does not invent hidden reasoning.

English and Chinese games share the same allowance with the dialogue game: 3 games per IP and 300 across the site each day, resetting at midnight Beijing time (UTC+8). Starting a game counts once; reloads and continued play do not. A shared network shares the allowance. Games and logs last up to 7 days; dealing again replaces the old game. Clues are sent to TypeSafe Jev, so leave out private information.

A game's language is fixed when dealt. Changing the page language does not translate an existing game's words or clues, or grant extra games.