Module 2.4.2 · Phase 2 Decision · Step 2.4 Metrics & ThresholdsPreferred
Robustness Testing
Do the fairness results hold under small changes and for every subgroup?
Tests whether your fairness results hold up under small perturbations and for intersectional subgroups.
What it does
On a dataset with model predictions and labels, you run sensitivity analysis, stress testing and a subgroup robustness audit. A checklist supports a sign-off for the testing documentation the EU AI Act asks for in Annex IV point 2(g).
Why it matters
This step translates fairness definitions into a binding measurement design, specifying metrics, thresholds, disaggregation requirements, and statistical validation criteria. It ensures that fairness evaluation is consistent, reproducible, and methodologically defensible.
Where it sits
- Fairness Definition
- Populations & Intersections
- Bias Identification & Prioritization
- Metrics & Thresholds
- Fairness Measurement Design
- Robustness Testing
- LLM Evidence Record
How to use it
In the Fairness Navigator Beta
Work through the steps in order. Several modules open from the Navigator's steps and hand their results back.
On its own Live
Open 2.4.2 from the platform's Modules page and run it when your project needs it. Work is saved with the assessment.
After a Pulse reading Live
Start with a fast first read, then go deeper with the modules where it points.