ACM Queue - AI Coding Benchmarks Need Composition Cards

My article AI Coding Benchmarks Need Composition Cards is live in ACM Queue. In June 2026, GPT-5.5 scored 70% on DeepSWE and 6.3% on the Diamond tier of FrontierCode. Neither number is wrong. The two benchmarks contain different work and grade it in different ways. I argue that every AI coding benchmark should publish a one-page composition card, with its task mix, its grader error rates, and a score for each category.