Position: We need theory-grounded explainability for abstract tasks to quantify model bias.