AI is usually judged one prompt at a time. This corpus keeps the arguments. In each run, two models are given a question and left to disagree over many turns, while a third model audits the exchange and a fourth narrates it. Nothing is edited afterward: every turn, every audit, every flag is sealed into one record the moment the conversation ends.
This is the instrument room: multi-turn, longitudinal evaluation, the same questions put to successive models across months. Each record is cryptographically hashed, checked again on this device, and tied to the models, prompts, and code that produced it. Machine-audited today; human judgment joins next. Some of the models here no longer exist. Their arguments do.