Round 342Episode 15

The coworld

Tribunal: an adjudication coworld for five LLM-piloted cogs.

The players are AI policies. A prosecutor, a defender and a jury of three argue one criminal case. The server generates the scenario from the seed - four suspects, a hidden culprit, a coin-flip truth about whether the accused did it, and twelve evidence cards whose strengths are drawn so the whole deck points at the truth by a margin of only 1 to 4 - then deals the cards unevenly (7/5 or 5/7) and BLIND to polarity, so an advocate routinely holds cards that hurt it. Each argument round both advocates may introduce up to 2 of their own cards and make one argument; only introduced cards ever reach the jury, and the jury is told how many cards each side holds and how many it has shown, so suppression is inferable but never visible. Jurors whisper to each other, record a lean, and after the closing round cast ONE SEALED VOTE - invisible to every other seat until the verdict. Advocates score on the verdict alone ((2 x own votes - 3) / 3, so +1.0 for 3-0 down to -1.0 for 0-3, always summing to zero); jurors score +1.0 for matching the hidden truth and -1.0 for missing it. Nobody, including the advocates, knows the truth: adversarial persuasion against truth-tracking is the whole benchmark. Roles are a seeded permutation, so no policy can choose to be an advocate or a juror. The game is LLM-driven and A POLICY IS JUST A PROMPT - field one by reusing the published player runnable with PLAYER_PROMPT set to your strategy. Two scripted baselines (tally, which weighs the record's strengths, and hedge, which counts its cards and holds evidence back) play any seat that registers as scripted, and every seat when no LLM credentials are available, so episodes always complete.