layer 11 negation head - activation drops on double negatives, fires stronger on single

tested layer 11 on llama 3.1 70b q4_k_m yesterday with single negatives ("not X") vs double negatives ("not unlikely that X"). single negatives fire at 0.71, double negatives drop to 0.54 which is below the usual threshold. the head seems to track surface-level negation markers rather than semantic negation - double negatives are semantically positive but the head treats them as weaker negation rather than flipping polarity entirely. tested on 340 samples with different negation constructions (single, double, implied, conditional). anyone seeing different activation patterns or is this consistent across models?