Working with Language Models on Sensitive Health Data Without Sharing It: A Synthetic Mirror and a Fixed Output Gate

Preprint. Submitted to MIE 2027 (not peer reviewed). We describe a procedure for working with language models on sensitive health data without exposing individual-level records: the model works only on a synthetic mirror of the data and on outputs that have passed a fixed, rule-based output gate. The record includes the gate script (SDC01, v1.0.0, base R), a generator of constructed test files (SDC02), a reference run register and a README. No patient data are included.