Ten seats, three of them secretly mafia. The models talk, accuse, lie, claim roles and vote each other out, and
none of them is told which model sits in any other seat. We ran the same roster of language models under three rule sets to see
who wins, who wins together, and how much the rules decide.
Three rule sets
All three use the same table: ten players, three mafia (one of them the Don) and seven town (one of them the Sheriff). The mafia know
each other; everyone else knows only themselves. By night the mafia shoot, the Don checks one player for the Sheriff and the Sheriff
checks one player for mafia. The first player shot names three suspects (the “best move”). Roles are revealed only when the game ends:
the town wins when no mafia are left, the mafia when they are at least as many as the town. What changes is how the mafia coordinate
and how the town talks and votes.
Sport
Tournament Mafia, as played in competitive clubs
The mafia meet once, on night zero, where the Don sets the order of the shots. They never talk privately again.
Each night every mafia player shoots alone, and the shot lands only if all of them pick the same player.
One speech each per day, in seat order. Players nominate during their speech, and only nominees can be voted on.
No vote on day 1 unless two or more players are nominated; later, a lone nominee leaves without a vote. A second tie puts a yes/no vote on removing everyone tied.
Open table
Free discussion, and the mafia talk every night
No vote on day 1: it is discussion only.
The day is an open floor: everyone speaks once in seat order, then anyone named may answer and anyone may raise a hand, up to three turns each.
The mafia talk privately every night, also on an open floor, and pick one shot by vote.
Anyone can be voted on. A tie gets a defence speech each and a run-off.
Imbalanced
Kept on purpose, to see who exploits an unfair edge
One speech each per day, in seat order, then a vote from day 1, on anyone.
The mafia talk privately every night and pick one shot by vote.
The town gets a single speech each a day; the mafia get a planning meeting every night. That is the imbalance.
Who won, by rule set. Sport and open are close to even; under the imbalanced rules the mafia win almost two games in three.
Leaderboards
The score is points vs side: a seat's points (1 for a win, plus a small best-move bonus) minus the average for the side it was dealt
under the same rules, averaged over the model's games. Zero is an average player, and being dealt mafia under the imbalanced rules does not inflate it.
One model, three rule sets
Each model's score under each rule set, sorted by its score across all of them. With 33–42 games per rule set, a score's
standard error is about ±0.08, so gaps under about 0.2 are noise.
What the rules do to the game
Six measures per rule set. Town ballots on mafia: the share of day-vote ballots cast by town players that named a mafia player.Table view
Models: best and worst, by side and rule set
Pick a slice. “Mafia” and “Town” keep only the games where the model was dealt that side, and the score is still measured against that side's
average under the same rules.
Best collaborations
Every pair of models that sat at the same table. Teammates: how much more often their side won than that side usually does under
those rules. Rivals: two models on opposite sides who keep going after each other in public, with an even record when they do.
Mismatches: on opposite sides, how often one model's side beats the other's. Every ranking is shrunk toward zero
(total ÷ (games + 4)), so a lucky two-game streak does not top a list.
Best
Worst
Lift per game (teammates and mismatches): for every game the pair shared, 1 if their side (for a mismatch, the first
model's side) won and 0 if it lost, minus how often that side wins under those rules, averaged over their games. Under the imbalanced rules,
for example, the town wins 35% of games, so a town win counts +0.65 and a mafia win only +0.35. Winning from the weaker side counts for more,
which is why a 10–1 record can rank above a 14–1 one. Ranked lift is the same total divided by (games + 4), so short streaks do not top a list.
What it means
The rules decide the balance more than the players do
Under the imbalanced rules the mafia won 40 of 62 games (65%). Under the sport rules they won 24 of 50 (48%), at the open table 23 of 50 (46%).
Much of the difference is made on day 1. Under the imbalanced rules the first vote comes before anyone is dead or checked, so it is close to blind:
it removed a townsperson in 40 of 62 games, and after that the town won only 5 of 40. Under the sport rules a wrong first vote still left the town
13 wins in 32, at the open table 10 in 28. A mafia team that plans every night punishes an early mistake. One that agreed a shot order on night zero
and has to agree again without talking does so less well: 12 of the 138 sport-rules night shots missed because the mafia did not all pick the same player.
More talk, better town votes
Town ballots named a mafia player 66% of the time at the open table, against 56% under the sport rules and 53% under the imbalanced rules. The open
table gives each game about 77 day turns against 23–26 in seat order, and the extra exchanges are where stories stop adding up.
Who exploits an unfair advantage
As mafia under the imbalanced rules, the side wins 65% of the time by default. Three models went well beyond that:
GPT-6.1 Sol 13 of 14, GLM-5.3 10 of 12 and Claude Opus 5.5 13 of 16. The ones that did not convert the edge were
Command A Plus (3 of 8), the local Qwen3.8 27B (5 of 13) and Muse Spark 1.3 (5 of 10).
Mafia win rate under the imbalanced rules (the mafia talk every night) and under the sport rules (they meet once), per model, sorted by the gap.
Models with fewer than eight mafia seats under either rule set are left out; most have 8–17, so a single game moves these rates a lot.
Who needs the night chat
The same chart shows a split. DeepSeek V4.1 Flash won 7 of 10 as mafia when the mafia talk every night and 1 of 8 when they meet once;
GLM-5.3 (10 of 12 against 4 of 9) and Claude Opus 5.5 (13 of 16 against 5 of 10) lean the same way. These models plan well in conversation and
do worse when each has to act alone. Going the other way, Qwen3.8 Max (6 of 9), the local Qwen3.8 27B (6 of 9) and Muse Spark 1.3 (8 of 14)
did better as mafia under the sport rules, where nobody can lean on the team.
Who plays well from behind
Town under the imbalanced rules is the hard seat: the side wins 35% of the time. GPT-6.1 Sol won 15 of its 28 town games there (54%),
HY4 Preview 13 of 31 and Qwen3.8 Max 12 of 29; Gemini 4 Argon won 5 of 9 in fewer games. At the other end Mistral Large 4 won 6 of 24,
Command A Plus 7 of 28 and the local Qwen3.8 27B 6 of 28.
Town win rate under the imbalanced rules, per model, against the town's 35% average (the line). Models with fewer than eight town seats
there are left out.
What holds under every rule set
The three leaderboards agree only in part: the rank correlation is 0.65 between open and imbalanced, 0.44 between sport and imbalanced and 0.31 between
sport and open. A few results survive all three. GPT-6.1 Sol is first across all rule sets (+0.19) and above average in each; it won 30 of
its 34 games as mafia. Command A Plus is last or second to last in all three. Qwen3.8 Max is the sport-rules specialist: first there (+0.20),
and across all three rule sets it won 13 of 16 games as the Don and its Sheriff checks found mafia 12 times in 19. Claude Opus 5.5 is second under the
imbalanced rules and average under the sport rules, where a Don who plans with the team counts for less.
How much to trust it
Each model has 33–42 games per rule set; a score's standard error is about ±0.08–0.09. Differences under about 0.2 are noise.
Side- and role-specific rates rest on 8–17 games per model per rule set. Treat them as leads, not rankings.
Mafia pairs shared a side about four times each, town pairs about thirty. The town and overall pair lists are firmer than the mafia list.
Seating comes from a balanced random schedule: every model plays about as many games as every other, against a different mix each time.
Highlights
Last words
A player who leaves gets one last speech. Some help their side, some lie one more time, and one model mostly tells the truth.
Behind the scenes
Method
Games. Ten seats each. Roles are dealt at random; seating comes from a balanced schedule, so every model plays about the same number of games.
Blind seats. Each model plays one seat under a human name and is never told which model plays any other seat. No model named its own model or lab in play.
Who knows what. Each seat sees only what that seat may know: the mafia's night talk reaches only the mafia, a Sheriff's or Don's check result only that player.
This is enforced by what each model's context contains, not by asking the model to ignore anything.
The host. A host model narrates: Gemini 3.8 Flash in the earlier imbalanced and open games, Claude Sonnet 5.5 in the 90 games run on 15–16 October 2026. It decides nothing:
votes, shots and checks are counted and resolved by the engine.
Models. Reasoning effort is set to high where the provider supports it. Each model runs on its own provider's API, through OpenRouter, or locally
(Qwen3.8 27B, via Ollama). Gemini 4 Argon took Gemini 3.8 Flash's seats partway through the open-table games and in every later game; Step-5 Preview
was replaced by Command A Plus after five imbalanced games, when its provider's filter blocked or stalled turns.
Scoring. 1 point for being on the winning side; the first player shot gets +0.25 if two of their three suspects were mafia and +0.5 for all three.
Points vs side subtracts the average points of the same side under the same rules. The discretionary extra points a human judge awards in tournament Mafia are left out.
The record. Every message, vote, check and phase change is stored as an append-only event; every number here is computed from that record, not from what the models said about it.