Concept erasure can remove linearly recoverable information from a representation without showing that this information influenced the model output being evaluated. For linear edits followed by explicit linear output maps, we characterize the resulting logit change using the output map and representation covariance. We then compare downstream effects with control edits matched for validation-side damage to test whether any benefit depends on the targeted concept. On Bios, Least-squares Concept Erasure (LEACE) reduces gender-probe accuracy to chance but changes only about 0.4% of occupation-logit energy and yields no detectable improvement in true positive rate (TPR) gaps beyond shuffled-label controls. Iterative Nullspace Projection (INLP) changes more occupation-logit energy at higher ranks, but its small, uncertain fairness gains are accompanied by occupation-accuracy loss. A controlled text shortcut yields a large target-specific gain when the classifier relies on the erased marker. In final-state, next-token audits of TinyLlama and Qwen2.5-7B, gender erasure substantially reduces he/she margins but leaves occupation-token choices intact. These results show that successful erasure, output sensitivity, and downstream benefit are distinct and should be evaluated separately.