Clinical text de-identification and audit assistant

Clinical notes cannot be shared for research until names, dates, IDs and contact details are removed, and tools make mistakes in both directions: they miss names that look like ordinary words and over-scrub medical terms. This is an open-source assistant for human-reviewed de-identification. Muse Spark finds identifiers, a second call checks the doubtful ones, and a replacement step shifts dates by one amount per patient and gives each real name one fake name. A red-team agent then reads the cleaned note and tries to work out who it is about. Every change links to its evidence (offsets, detector, confidence, verifier verdict), and anything left undecided stays replaced. It is an assistant for human review, not a compliance certification. All data is synthetic.