Split Decision is a solo-built project, an honourable mention in the Global AI Hackathon Series with Qwen Cloud, that stages Supreme Court deliberations with nine AI judges holding different judicial philosophies. A clerk agent reads case data from CourtListener, Oyez and the Supreme Court Database and writes a bench memo; nine jurors on qwen3.7-plus cast secret ballots, argue in rounds, debate in pairs chosen by a foreperson, and revote until a 7-2 supermajority appears or round five is reached. The anti-sycophancy design is explicit: private votes before anyone speaks, personas re-injected every round so judges do not drift toward agreement, and written rules governing when a judge is allowed to change its vote. Every ballot, flip and verdict is written to one schema-validated event log, and that log drives a scoreboard that grades the panel against the Court's real decisions across four conditions - solo model, silent jury, full debate and a neutral control. Only cases decided after the model's training cutoff are scored, and predictions on pending cases are put on the record so later outcomes grade them. The reported result is 66.7% in both conditions across 24 paired cases: debate made the panel more articulate, not more accurate. Qwen Voice Design produced twelve voices, wan2.6-t2i generated episode art, and the pipeline runs on one Alibaba Cloud Simple Application Server with OSS holding the audio and feed.
Our takeThis is one of the few public attempts to measure multi-agent debate rather than assume it, and the null result is the part worth believing. The scaffolding deserves copying more than the verdict: private first ballots, personas re-injected each round, explicit rules for when a vote may change. Twenty-four paired cases is a small sample, so the honest reading is no demonstrated gain, not no gain at all.