Libra: A Prototype for Provenance-Aware Fairness Audit Parameters, Implementation Review, and Descriptive Sensitivity Analysis

Abstract

Fairness audits require choices about groups, reference populations, outcome definitions, and decision thresholds. This unreviewed technical preprint presents Libra, a prototype that stores some of those choices with process metadata and applies a selected threshold and reference group in a group-outcome audit. It reviews implementation boundaries and reports descriptive sensitivity analyses using one 2024 Michigan HMDA lender extract and a historical ProPublica COMPAS extract. The HMDA Black-to-White origination ratio is 0.8174 across retained action codes and 0.7162 for originated-versus-denied records under the stated filters; these are different administrative-outcome estimands, not AI-model performance measures. The paper uses no community participant data, validates no community-defined standard, and establishes no government endorsement or adoption. The public companion package documents source acquisition, scripts, results, and their limitations.