Fourteen papers covering eval methodology and production observability. SPADE is the assertion-mining pattern that ties prompt versions to invariants. RewardBench, MixEval, and WildBench are the leaderboard and cost-correlation results to read before adopting any meta-benchmark. IFEval, InfoBench, and FollowBench are the instruction-following trilogy: atomic verifiable constraints, DRFR, and stacked difficulty. The agent-eval survey from Yehudai et al., IBM's observability paper, and the data-centric position paper are the references for the production end of the pipeline. ScienceAgentBench and SWE-bench Multimodal are pulled in here because their findings are about eval methodology as much as agent capability. The 2026 pair is about comparisons that mislead. Tran and Kiela show that many multi-agent gains disappear once reasoning tokens are held constant, and the HORIZON paper shows how to attribute long-horizon agent failures with a judge that has been checked against human labels.