Show HN: A Replayable A2A Jury That Traces Whether One Agent Actually Changed Another's Mind — and Refuses to Claim Causation
The ai_courtroom example in nMaroulis/protolink (88 stars, Python) simulates a liability tribunal over an autonomous-vehicle fatality with seven interacting pieces of evidence, using agent-to-agent messaging where every juror returns an observable public action carrying its message, target, evidence citations, and explicit intent. Private guilt probabilities are never auto-exposed to peers — only what an agent chooses to reveal — and the replay records before/after probability shifts and vote flips tied to specific exchanges. The intellectually honest detail worth stealing: the docs label these 'observed after-message shift' in explicitly non-causal language, because correlation doesn't prove influence without ablation. It compares four jury topologies (solo, independent, star foreperson-hub, full mesh) to isolate how communication structure alone changes group decisions.
↳ Follow the thread