Fine-grained sexism classifiers are known to rely on surface gender terms, and explainability methods are increasingly proposed not merely to diagnose this reliance but to correct it during training. We test that proposition directly. Using LIME, we identify the tokens a DeBERTa-v3-large classifier relies on most when it misclassifies EDOS Task C development posts, and penalise those tokens’ hidden-state magnitudes during continued fine-tuning. Across three random seeds the method does not improve macro-F1: the best penalised setting reaches 0.477 against a 0.468 baseline (within seed variance), and continued training with the penalty disabled scores highest of all (0.483). A controlled ablation isolates why. Suppressing only the gendered subset of flagged tokens, only the function-word subset, and a random frequency-matched set all leave test macro-F1 statistically indistinguishable from the baseline (0.466, 0.467, and 0.461); critically, LIME-guided suppression performs no better than random suppression, indicating the attribution signal carries no usable corrective information. The flagged tokens are themselves predominantly gendered terms and slurs (65–80% of LIME’s top tokens across seeds) — that is, the genuine evidence for sexism rather than spurious cues. We further find that attribution methods disagree sharply on which tokens are most influential (LIME 65%, SHAP 30%, Integrated Gradients 35% gendered), cautioning against single-attributor conclusions. We conclude that attribution magnitude is an unreliable criterion for token suppression in abuse detection, and that explanation-guided correction cannot succeed where the most salient features are the task’s real signal.