← Back to article | Measuring the gap: correlating synthetic-to-real drift with PHI de-identification performance