Corrected Without Looking: Post-Hoc Long-Tail Rebalancing Leaves Model Evidence Provably Unchanged

Our contributions are:

A proof that τ-normalisation, LWS, and post-hoc logit adjustment are exactly Grad-CAM invariant, verified over 47,952 comparisons (Sections 3 and 4).
The Explanation Migration Index (EMI), which separates target switching from evidence change and is analytically calibrated on the invariant family.
Evidence that attention rollout is bit-identical under every head operation, helping isolate where explanation change originates.
A quantification of lesion-level leakage in the HAM10000 dataset that standard duplicate-detection methods cannot identify.