Stable attribution maps may suggest that a medical AI model relies on intended image features, but stability alone cannot show which factor drives the prediction. We examine this gap in attribution-based medical XAI. Audited decision dependence is prediction-side dependence on a pre-registered target factor, tested by admissible targeted interventions against matched controls under the same audit budget on a pre-registered model-output response scale; label flips and continuous probability or logit changes are explicit instances. Building on the established distinction between observational stability and intervention-based dependence, we show that target-dependent and shortcut-dependent predictors can be indistinguishable under non-target transformations yet respond differently when the declared factor is modified. We introduce audited evidential dependence as a stronger framework requiring graded sensitivity, refutability, and distribution-level prevalence, with a minimal audit protocol using admissibility checks and uncertainty-aware verdicts. A controlled BraTS2020 tumour-mask versus shortcut experiment provides a quantitative proof-of-concept of the audit logic, not end-to-end attribution or clinical validation. The primary audit uses official tumour masks as supplied candidate regions and constructed target-oriented and shortcut-oriented control predictors, testing a designed dependence contrast rather than attribution-based discovery of a clinically meaningful target. Both predictors remained prediction- and attribution-stable under mild non-target perturbations, whereas the matched-control audit showed a large positive control-subtracted tumour-factor response for TargetModel and a near-zero response for ShortcutModel. Deletion-style sensitivity was complementary but not target-specific. The framework provides a practical audit layer for testing whether a stable or plausible explanation reflects dependence on the intended target factor rather than a correlated shortcut.