For every concept word, the paired difference in concept-vector cosine against the no-mention baseline at the selected residual layer: dot = mean over that word’s sentences (token-mean per sentence), whisker = ±1 std across sentences.