Post-processing mitigations modify the model’s outputs after inference, avoiding model retraining. They are the simplest to implement but may be criticised as cosmetic corrections that mask underlying model bias without addressing root causes.

Threshold calibration adjusts the decision threshold for each subgroup to equalise selection rates or error rates. Fairlearn’s ThresholdOptimizer automates this, finding per-subgroup thresholds that satisfy a given fairness constraint while maximising accuracy. The AISDP documents the per-subgroup thresholds, the fairness constraint targeted, and the resulting impact on the overall accuracy and per-subgroup error rates.

Score adjustment applies additive or multiplicative corrections to the model’s raw outputs for specific subgroups. Calibrated equalised odds (Pleiss et al., 2017) adjusts probability scores per subgroup to achieve both calibration and equalised odds simultaneously. The adjustment parameters are documented along with the mathematical formulation.

Reject option classification routes borderline predictions (where the model’s confidence is low) to human review, reducing the chance that uncertain predictions disproportionately harm one subgroup. The confidence threshold for routing to human review is documented, along with the expected volume of cases routed and the capacity of the human review process to handle them.

The AISDP must document why post-processing was chosen over root-cause mitigation. Valid reasons include that the root-cause mitigation would require protected characteristic data the organisation cannot lawfully obtain, that root-cause mitigation would reduce accuracy below declared thresholds, or that the bias is an artefact of historical data that cannot be corrected within the training data. The AI Governance Lead signs off on the choice.

Key outputs

• Post-processing technique selection and rationale
• Adjusted parameters (thresholds, scores) per subgroup
• AI Governance Lead sign-off on technique choice