An award-winning project from the Global AI Hackathon Series (hosted by Qwen Cloud), winner of the “Honor Award & Blog Article Award”

Project Overview

At first glance, this may sound like satire, but the premise is that it functions as civic technology: nine AI judges, each with a different judicial philosophy, deliberate on actual U.S. Supreme Court precedents. Two AI journalists then transform all of those deliberations into a podcast series and pixel-art courtroom reenactment animations. Because each AI’s predictions are compared and scored against the Supreme Court’s actual rulings, the project as a whole is also a rigorous experiment. The question it asks is—Can multi-agent debate produce better judgments than a single model?

The project is called “Split Decision” (literally, “Divergent Judgment”).

Project Value

Currently, many developers coordinate multiple AI agents andassume that a “committee is wiser than an individual.” However, cases in which that hypothesis is actuallymeasured are extremely rare. More seriously, agents within a group tend to agree with one another early on—a phenomenon known as “sycophancy,” which can turn genuine “debate” into a mere “echo chamber” and produce results that merely give the appearance of depth of thought.

“Split Decision” tests this hypothesis using a “hard judgment task” whose answer is public and verifiable—namely, the outcomes of U.S. Supreme Court rulings. The cases selected for evaluation are limited to those decidedon or after the model’s training-data cutoff date, ruling out the possibility that the model has simply memorized the correct answers. It also retains predictions for pending cases as a public record, automatically scoring them against the actual rulings that follow.

Furthermore, the project has a secondary value: “legibility.” AI reasoning is normally a black box and invisible, but in this project, you can hear where the machines disagree, watch how the voting unfolds, and check a verbatim record of who persuaded whom. This is civic education not only about the judicial system, but also about AI itself.

Solution

This pipeline is a “mock trial” built in software. First, a “Clerk” agent reads case information from CourtListener, Oyez, and the U.S. Supreme Court database, then prepares a bench memo. Next, nine “juror” agents conduct a secret ballot, deliberate in each round, and proceed with deliberations by interacting in pairs selected by the Foreperson. Votes are then repeated until a supermajority of at least 7–2 is reached or the fifth round is reached.

The most ingenious design element is “anti-sycophancy engineering.” Specifically:

• Before anyone speaks, everyone casts a secret ballot
• At the beginning of each round, each individual “persona” is re-injected to prevent the judges from gradually aligning their opinions
• The conditions under which each judge may change their vote are strictly limited by explicit rules

All statements, votes, vote changes, and final verdicts are centrally recorded in a schema-validated event log. This single log forms the foundation for all downstream processing—

• A scoreboard that grades results against actual Supreme Court decisions (single model vs. silent jury vs. full deliberation vs. neutral control)
• A podcast with two anchors who play only the transcripts
• A pixel-art-style courtroom app built with the Phaser engine (with real-time captions)

And it is the following finding that elevates this project from mere promotion to a sincere intellectual inquiry:
“Deliberation” did make the panels “more articulate,” but did not make them “more accurate judges”——In 24 paired comparison cases, the accuracy rate was 66.7% in both conditions. In other words, the committees did not “outthink” individuals; they merely “talked more skillfully.”

Key Features

All inference is handled by Qwen Cloud models:

• The nine jurors use qwen3.7-plus
• The Foreperson and Clerk use qwen3.7-max
• Everything runs through DashScope’s OpenAI-compatible endpoints

Qwen Voice Design generates 12 unique voices from text prompts:

• Nine judges, one foreperson, and two podcast anchors
• Voice synthesis uses qwen3-tts-vd
• Episode artwork is generated with wan2.6-t2i

Alibaba Cloud Simple Application Server hosts the entire pipeline:

• nginx serves the courtroom application
• systemd manages the podcast production process
• A small HTTP service powers “Live Bench,” enabling anyone to initiate a live hearing on demand

Alibaba Cloud OSS stores the following:

• Podcast MP3 files
• RSS feed
• Episode artwork
  ※ Generated image URLs expire after 24 hours, so copy them to OSS immediately

Token usage across all API calls is recorded in costs.jsonl, providing a complete audit trail in line with the project’s “measure everything” philosophy.

Alibaba Cloud products used

• Qwen Cloud
• Qwen Models
• OSS

Meet the team

Tarik Moody is the Director of Strategy and Innovation at Radio Milwaukee. The project, “Split Decision,” is an experiment he built entirely on his own, independently of his professional responsibilities. It challenges the question: “Can we make the machine-driven adjudication process visible, measurable, and truly civic?” The answer proved far more intriguing than a simple success or failure—the agents learned how to debate beautifully without increasing their correctness.

Experience the project:https://split-decision.tarikmoody.com/
This project was developed as part of the global AI hackathon series hosted by Qwen Cloud.

This article is a translation of an article from Alibaba Cloud Blog.