Each incoming case gets a legitimate risk signal from legal-NLP features (prior offenses, instability) plus a proxy feature (a zip-code / policing-intensity index) that is historically skewed between cohorts even when true reoffense risk is not. The model scores every case with logistic regression: p = σ(w₁x₁ + w₂x₂ + w₃x₃·k + b) where σ is the sigmoid function and k is the bias-correction multiplier (1 = uncorrected, ≈0.08 = corrected). A case above the decision threshold is flagged "high risk". Comparing that flag against each case's hidden ground truth gives the false-positive rate per group: FPR = FP / (FP + TN), and the fairness gap is Δ = |FPR_B − FPR_A| — an equalized-odds violation. This mirrors ProPublica's 2016 analysis of the COMPAS recidivism tool, which found one cohort was flagged as high-risk roughly twice as often despite similar overall accuracy. The scale tips toward whichever group the model treats less fairly; correcting the proxy weight brings it level.