Fairness and robustness gates in CI/CD

WHAT THIS TESTS Whether you can translate fairness and robustness from concepts into automated, blocking checks inside a deployment pipeline.

A GOOD ANSWER COVERS Treat fairness and robustness as test stages that run on every candidate model before promotion, just like unit tests gate code. For fairness, evaluate on a curated held-out set sliced by protected attributes and other important subgroups; compute metrics such as per-slice accuracy and the gap between groups, demographic parity difference, and equalized odds difference. For robustness, run the model against perturbed inputs (noise, typos, occlusions, image corruptions), distribution-shifted and adversarial test sets, and check that performance does not collapse beyond a tolerance. Encode predefined constraints as numeric thresholds in config, for example maximum allowed accuracy gap between groups and minimum accuracy under perturbation.

HOW TO FAIL THE BUILD The evaluation stage loads the candidate model, runs the metric suite, compares each result to its threshold, and exits with a nonzero status if any constraint is violated; the CI orchestrator then blocks the deploy stage. Emit a machine-readable report and a human-readable artifact so reviewers see which slice or perturbation failed. Store results per model version for trend tracking.

COMMON WRONG ANSWERS Gating only on overall accuracy, hiding subgroup regressions; running fairness checks manually and inconsistently; or logging violations as warnings that do not block promotion.

LIKELY FOLLOW-UPS How to choose thresholds, handle small subgroup sample sizes, and avoid the fairness suite leaking into training.

ONE CONCRETE EXAMPLE A loan model passes overall accuracy but the pipeline computes a twelve point accuracy gap between two demographic slices, exceeding the configured five point limit. The fairness stage exits nonzero, the build fails, deployment is blocked, and the report flags the failing slice for the team to investigate.