A colleague wants to replace dropout positives with LLM...

While genuine positives often beat dropout, the risk in clinical notes is high. You must audit a stratified sample of the generated paraphrases for errors in negation, dosage, and laterality, where mistakes can invert the clinical meaning. Additionally, consider data residency constraints and the fact that generated pairs inherit the teacher model’s blind spots. The final decision should be based on a performance comparison on a mined-pair evaluation set after the audit passes.