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