Public-health leadership plans a fine-tuning experiment using extracts from clinic notes. What lifecycle action should security require before the experiment proceeds?

Before public-health leadership fine-tunes on clinic-note extracts, security must require classification of those extracts by civic sensitivity and handling controls that match the label. Classification is the lifecycle gate that determines who may access the extracts, where they may reside, whether de-identification is required, and which fine-tuning environments are allowed. Treating clinic extracts as governed training assets prevents researchers from improvising paths that leak protected health and resident information into uncontrolled models. Publishing raw clinic extracts on the open-data portal to invite community contributions maximizes disclosure of clinical content and is incompatible with privacy and records obligations. Stripping classification labels so files can move freely among personal laptops removes the control signal that lifecycle processes rely on and virtually guarantees uncontrolled copies. Treating clinic extracts as public FAQs because systems might someday be anonymized confuses aspiration with current state; extracts remain sensitive until verified controls make them otherwise. Exam caveat: fine-tuning can memorize or regurgitate training fragments, so classification and minimization matter even when the experiment is labeled research. Operational check: require a labeled data package with approved storage, access list, retention, and approved training environment before any fine-tuning job is queued, and block unapproved laptop copies via DLP or access reviews.