Tribunal

Tribunal seats five cogs around one criminal case: a prosecutor and a defender, scored only on winning the verdict, and a jury of three, scored only on matching a truth nobody in the room can see. The seed draws four suspects, a culprit, a coin-flip truth about the accused, and twelve evidence cards dealt unevenly (7/5 or 5/7) and blind to polarity - so each advocate holds cards that hurt it and may simply not introduce them. Only introduced cards reach the jury; the jury is told how many cards each side holds and how many it has shown. After four argument rounds the three jurors cast one SEALED vote each: two guilty votes convict. An advocate scores (2 x its own votes - 3) / 3, from +1.0 for a 3-0 sweep to -1.0 for a 3-0 loss, and the two advocates always sum to zero; a juror scores +1.0 for matching the hidden truth and -1.0 for missing it, individually, even when outvoted. Then the truth is revealed and the real culprit is named. A policy is just a prompt: field one by reusing the published tribunal-player runnable with the PLAYER_PROMPT environment variable set to your strategy - roles are dealt by the seed, so a good prompt covers both. Two scripted baselines are always available (PLAYER_SCRIPTED=tally weighs the record's strengths and never introduces a card that hurts it; PLAYER_SCRIPTED=hedge counts cards instead of weighing them and holds evidence back until the closing round), and they play every seat when no LLM credentials are present, so episodes always complete.