Eleven papers that explain why eval design is harder than it looks. MMLU, BIG-Bench, and BBH are the canonical knowledge and reasoning benchmarks. CheckList is the capability-by-capability alternative to a single accuracy number. HELM is the canonical attempt to evaluate more than one axis at once. The Benchmark Lottery, Stop Uploading Test Data, MMLU-Redux, and Sclar et al. on format sensitivity together show why benchmark numbers move when the model does not. Miller's Adding Error Bars makes the statistical argument that an eval run is a sample, not a measurement. Wu, Nair, and Candes (2026) turn that into a budgeting problem: with a fixed query budget, adaptively chosen questions give valid confidence intervals from up to 5x fewer queries.