What Transfers and What Does Not: A Register-Level Diagnosis of Synthetic-to-Real Transfer in Clinical PHI Detection

Synthetictrainingdatahasbecomeastandardresponsetothedifficulty of obtaining real clinical text, and the distribution shift it induces is by now documented. What has not been characterised is that the shift is not uniform within a clinical document. The detector we examine reaches 93.0% F1 on its own synthetic test data. Applied to MIMIC without adaptation, on the same category and within the same note, it reaches 72.2% precision on content written as formal diagnosis and a recall of 0.001 on the telegraphic, abbreviated, frequently negated shorthand of the social history section. The two are measured over different regions: recall inside the social-history region against a reference annotation, precision outside it by manual adjudication, with recall on the formal register not measured because no reference annotates it. The system is a two-stage classifier-plus-NER detector for GDPR Art. 9 special-category content, as distinct from the HIPAA Safe Harbor identifiers that de- identification removes while preserving exactly the content Art. 9 regulates. Two further diagnoses follow. An annotation-scope confound affects any reuse of a phenotyping corpus to evaluate detection: scored naively over whole notes such a reference reports near-total failure for predictions that are largely correct but fall in unannotated regions, costing here a factor of seven in reported precision. A span-fragmentation failure traces to a training-data construction error invisible at training time. We quantify the operational cost and release the evaluation harness.