arXiv paper: KAISEN: Reproducible Subgroup Fairness Auditing for Clinical Risk Models

An arXiv paper introduces KAISEN, a five-phase pipeline for auditing subgroup fairness in clinical risk models, and stress-tests it to failure across 16 synthetic disease tasks and 15 social-determinant axes. The work supplies four empirical findings on the reliability of common audit components: significance measures are sensitive to minimum detectable effect floors, post-hoc mitigation methods show counterintuitive variability, mechanism diagnostics fail silently under proxy misspecification, and drift monitoring thresholds do not transfer across cohort realizations. The authors release all code, artifacts, and scripts to reproduce every reported number, emphasizing that the results are synthetic with known ground truth and carry no clinical validity claims.