Certified Fairness Auditing of Artificial Intelligence Flood-Risk Zoning Under Historical Redlining Legacies: Disparity Bounds and Insurance Access Equity

The central problem is that standard group-fairness metrics treat protected attributes as fixed, whereas in formerly redlined districts the relevant disparity is mediated by structural disinvestment that is itself a function of past policy. The paper formalises this setting as a constrained risk-classification problem in which the auditor seeks distribution-free upper bounds on the disparity between approval rates for flood insurance across historically redlined and non-redlined tracts, conditional on a small number of empirical moments. Three contributions follow. First, a definition of legacy-conditional demographic parity is introduced that separates the direct effect of a protected attribute from the indirect effect of redlining-linked covariates. Second, sharp bounds on the disparity are derived using linear programming over the space of joint distributions consistent with observed marginals and a bounded total-variation constraint; the bounds are certified in the sense that they hold for every distribution in the ambiguity set. Third, the framework is applied to a stylised flood-risk zoning pipeline calibrated to the Greater Harare metropolitan area, using NASA POWER monthly reanalysis air temperature at 2 m for Harare, Zimbabwe (-17.83, 31.05) over 2000-2023 (312 monthly observations; mean 18.914 C; warming trend 0.2001 C per decade). The analysis is methodological: it does not report empirical disparity estimates, but it shows how the certified bounds would be computed and interpreted, and it identifies the conditions under which a zoning model can be certified as legacy-fair. Certification should be a precondition for public deployment of AI flood-risk zoning, and that the bound width itself is a policy-relevant quantity.