fairness-testing

Fairness Testing

Fairness is not a property you check once — it's a property you test continuously. These tests run in CI and fail the build if the model regresses on fairness metrics.

The RED-GREEN-REFACTOR Cycle for Fairness

RED: Write a fairness test. Run it. Watch it fail (or confirm the threshold matters). GREEN: Adjust the model or data until the test passes. REFACTOR: Clean up. Do not delete the test.

This is non-negotiable. If you write no fairness tests, you have no fairness guarantees.

---

Test Structure

Every fairness test follows this pattern:

def test_<metric>_parity_across_<attribute>(model, test_data):
    """
    <metric> must not differ by more than <threshold> across <attribute> groups.
    
    Regulatory basis: Equality Act 2010 s.149 (PSED) / EU AI Act Art. 10
    """
    results = {}
    for group in test_data[attribute].unique():
        mask = test_data[attribute] == group
        results[group] = compute_metric(
            model, test_data[mask]
        )
    
    max_gap = max(results.values()) - min(results.values())
    assert max_gap <= THRESHOLD, (
        f"{metric} gap across {attribute} is {max_gap:.3f} "
        f"(max allowed: {THRESHOLD}). "
        f"Group breakdown: {results}"
    )

---

Required Test Suite

Write ALL of these. Skip none.

Test 1: Demographic Parity

import pytest
import numpy as np

DEMOGRAPHIC_PARITY_THRESHOLD = 0.05

def test_demographic_parity(model, test_df):
    """Positive prediction rate must not vary by more than 5% across protected groups."""
    protected_attrs = ['sex', 'ethnicity', 'age_group']  # adjust to your attributes
    
    for attr in protected_attrs:
        if attr not in test_df.columns:
            continue
        
        rates = {}
        for group in test_df[attr].unique():
            mask = test_df[attr] == group
            if mask.sum() < 30:  # skip groups too small to be meaningful
                continue
            X = test_df[mask].drop(columns=['label'])
            rates[group] = model.predict(X).mean()
        
        if len(rates) < 2:
            continue
        
        gap = max(rates.values()) - min(rates.values())
        assert gap <= DEMOGRAPHIC_PARITY_THRESHOLD, (
            f"Demographic parity violation on '{attr}': gap={gap:.3f} "
            f"(threshold={DEMOGRAPHIC_PARITY_THRESHOLD}). "
            f"Rates: {rates}"
        )