Counterfactual Marginalisation for Robustness Testing

This framework evaluates model robustness by using counterfactual image generators to intervene on nuisance variables like age or sex during test-time. By averaging predictions over these interventions, the method produces stability and calibration metrics that account for demographic shortcuts.