The Diagnostic Audit: Medical AI You Can Finally Argue With

Medicine is learning that a diagnosis a clinician cannot argue with is one the clinician cannot use. Medical-image AI today offers two doors: accept the model's call or reject it. A new arXiv preprint proposes a third — ask the model to write a six-part receipt for its reasoning, then audit it line by line. Call it the diagnostic audit.

The mechanism is borrowed from a 1958 model of argumentation, split into six named parts: claim, grounds, warrant, qualifier, rebuttal, backing. The preprint maps each role to a check. A biomarker-extraction model supplies the grounds — the visible signs in the scan. A medical-knowledge language model (MedGemma) reviews the warrant — whether the evidence licenses the conclusion. An image-similarity model (MedSigLip) builds the rebuttal, flagging scans unlike the training set. A qualifier grades confidence. A human expert reads the whole thing and decides.

The pattern generalizes. Any high-stakes AI call can be turned from oracle into audit by mapping its reasoning onto a named argumentation schema, assigning each part a check, and reserving the final call for a person who can read the receipt. The clinician's job does not shrink; it moves up from verifier to auditor.

The limit is real. This is a single arXiv preprint, no peer review, no clinical trial, and the audit is only as honest as the parts that fill it. The win is not a smarter oracle. It is a medical AI you can finally argue with.