Fairness After the Fact: Reporting Lag and the Blind Spot in Governance Metrics

Models certified as fair at deployment degrade as distributions shift, and temporal fairness frameworks respond by monitoring deployment, accumulating harm, and correcting once bias crosses a threshold. These frameworks assume violations can be observed as they occur. In high-stakes settings they cannot: loan default, recidivism, and hiring success take months or years to confirm, so evidence of bias arrives long after the decisions that caused it. We show the consequence is not merely late repair but that current metrics cannot distinguish a responsive system from an unresponsive one, since they measure latency from detection rather than from harm. We decompose repair latency as $M_E = M_A + \Delta_r$, separating an algorithmic component from a structural floor set by the reporting pipeline, and report their sum as Effective Mean Time to Fairness Repair. On COMPAS under measurement-bias drift, increasing lag reduces intervention frequency and drives disparate impact below the regulatory threshold with model, drift, and correction mechanism held fixed. The penalty is uneven: because debt is weighted by the number of individuals affected, a fixed global threshold produces a detection cliff rather than a gradient, and the smallest monitored subgroup never triggers intervention under any condition despite sustained violation. A per-capita threshold restores detection at every size.