LLM-as-judge, RAGAS/DeepEval, golden datasets, agent trajectory evals and CI regression suites.
Evals are the executable spec for a system whose behavior you cannot reproduce by reading the code — this is how you build, gate, and monitor that spec.
A fraud detector that flags nothing scores 98% accuracy and catches zero fraud — this lesson is the arithmetic that stops you from shipping that.
A model grading another model is a measurement instrument with systematic biases — calibrate it against humans, or you are optimizing noise.
Decompose RAG and agents into the components that can break independently, score each one, and gate every change so a regression points at the broken layer instead of a vibe.
Turn evals into a merge gate that blocks regressions, controls judge spend, defends against contamination, and pages you when production quality drifts.