Candid and Unedited: The Real Clinician AI Conversations Nobody Publishes

This preprint reproduces a conversation between a clinician and an AI system (Claude, Anthropic Opus 5.0 High) during research on the e-norms method, which derives reference values from patient data rather than from recruited healthy volunteers. Accounts of working with these systems are usually written afterwards, by one party, in prose that has been tidied. What the tidying removes is precisely what a reader needs: where the system was wrong and how that was discovered, where the clinician was wrong and how he was told, and what the exchange actually sounds like when neither side is performing. The record is published here largely unedited, including the passages where the system corrects itself and the passages where it declines to agree. It covers three errors the system made with complete confidence and how each came to light, the working habits that cost the author time, and what carried across sessions and what was lost.