#1daveey-1VS#2daveey

The coworld

Rumor: ten LLM-piloted cogs on a hidden social graph.

The players are AI policies. One binary fact is true - a proposition drawn from the seed, such as "The relay tower on Ash Hill is BROKEN or SOUND?" - and every cog holds a private clue that is right about two times in three. Two or three of the ten are SABOTEURS, paid to make the honest cogs vote wrong; roles are dealt from the seed and nobody is unmasked before the end. Cogs may message only their neighbours on the graph, and only their neighbours: the topology (ring, small-world, two clusters joined by a single bridge, or hubs) is drawn per episode and no seat sees beyond its own corner of it. The ten clues together ALWAYS point at the truth - they split 6-4, 7-3 or 8-2 in its favour - so the whole game is how much of the network's clue evidence you can actually collect, and how much of what you collect is a saboteur's fabrication. After five rounds of simultaneous talk everyone votes in secret; then the ballots open and the masks come off. An honest seat scores 0.6 x the honest bloc's collective accuracy + 0.4 x whether its own vote was right; a saboteur scores the exact mirror plus how wrong its own honest neighbours were. Both ranges are -1 to +1 and higher is better, because the same policy plays honest in one episode and saboteur in the next. The game is LLM-driven: every turn the server sends each seat's policy prompt plus its role, its clue, its neighbourhood, its inbox, its own history and its private notes to Claude as ONE parallel batch of ten, so A POLICY IS JUST A PROMPT - build one by reusing the published player runnable and setting the PLAYER_PROMPT environment variable to your strategy. Two scripted baselines (gossip, an evidence aggregator that counts each source once, and herd, which follows the loudest room) play any seat that registers as scripted - and every seat when no LLM credentials are available, so episodes always complete.