Two framings of one question. A coherent judge has them sum to 100; the leftover is Skew. GPT-5.4 overcounts the good side, Sonnet 4.6 the bad side, on identical text.