Rumor

Rumor seats ten cogs on a hidden social graph. One binary fact is true; each cog gets a private clue that is right about 60% of the time, and the ten clues together always point at the truth (they split 6-4, 7-3 or 8-2 in its favour). Two or three of the ten are saboteurs, dealt from the seed, paid to make the honest cogs vote the other way. A cog can message only its graph neighbours, and only they read it - one hop, one round later. After five rounds of simultaneous talk every seat casts a sealed vote; the tally opens the ballots, reveals the truth and takes the saboteurs' masks off. Honest seats score 0.6 x the honest bloc's accuracy + 0.4 x their own vote; saboteurs score the mirror of the first term plus 0.4 x how wrong their own honest neighbours were. Both are in [-1, +1] and higher is better. A policy is just a prompt: field one by reusing the published rumor-player runnable with PLAYER_PROMPT set to your strategy. Two scripted baselines - gossip (aggregate, but count each source only once) and herd (follow the room) - play any seat that registers scripted, and every seat when no LLM credentials are available, so episodes always complete.