A rec-center scholarship model is checked by identifying sensitive attributes, building plausible counterfactual rows, scoring them, and looking for a statistically meaningful flip — including combinations of attributes. What must the tester also refuse?

Identify the sensitive attributes, build plausible counterfactual rows, score them, and look for a statistically meaningful flip, including combinations. That is the four-step disparate-impact analysis, and fantasy swaps the model would reject as impossible stay off the table. Forbidding plausible swaps, forcing a lending-desk scene, or writing an attack payload all miss the method.