AIMS Proof-of-Concept: Reproducible companion code and audit log for Section 5.5 of "Legitimacy Conversion as Organizational Infrastructure: A Three-Stage Feedback Mechanism for AI Governance"

Deterministic Python implementation (seed 42) of three AIMS governance controls (Pillar 2 equity evaluation and proxy removal; Pillar 4 SHAP-based local audit; Pillar 5 threshold-recalibration sweep) operationalized on the UCI Heart Failure Clinical Records dataset (Chicco & Jurman, 2020). The archive contains the script poc_aims.py, the audit log produced by the canonical run (aims_audit_log.jsonl), and the metadata required to reproduce every numerical value reported in Section 5.5 of the manuscript.\n\nHeadline empirical finding: removing the dataset's strongest clinical predictor (serum_creatinine) as a candidate socioeconomic proxy *increased* the intergroup AUC gap (0.013 -> 0.043) rather than reducing it, while only modestly degrading overall AUC (0.910 -> 0.896). This counter-intuitive result, consistent with the fairness-accuracy trade-off literature (Ktena et al., Nature Medicine 2024), is the kind of contested technical signal for which the paper's IFP Stage 2 interpretive mediation is designed.\n\nThe underlying dataset is downloaded automatically by the script from the UCI Machine Learning Repository on first run; it is not redistributed in this archive.\n\nSee README.md for full reproduction instructions and expected output values.