Evals or it didn't happen
Somewhere right now an engineer is telling a stakeholder that the new prompt is better. Asked how they know, they will say it looks better. That sentence is where AI systems go to rot. The alternative is a small machine, cheap to build, that converts claims about AI behavior into measurements: golden sets, judges with named biases, and gates that fail builds.
Golden sets, built honestly
The foundation is a versioned collection of cases with known good answers or graded rubrics. Honest means three populations: realistic cases sampled from actual usage, hard cases where the system historically struggled, and adversarial cases designed to hurt, injection attempts, absurd inputs, questions with no answer in the corpus. Twenty of the first, five of each of the others is a real starting point. Keep it in the repo, change it through review, and treat it like the test suite it is. A golden set that only contains easy wins is a mirror for flattering yourself, not an instrument.
Judges, with their biases named
For outputs a script cannot score, use a model as the judge, but only with its known biases mitigated. Position bias: judges favor whichever answer they saw first, so run every comparison both ways and demand agreement. Verbosity bias: longer answers score better independent of quality, so normalize or penalize length. Self-preference: models favor their own writing style, so judge with a different model than the one that generated. A judge without these three mitigations is a random number generator wearing a robe.
Three altitudes
Evals live at three heights, and each catches what the others cannot. Offline experiments, run by hand while iterating: fast, exploratory, disposable. Gates in CI, run on every change: a fixed suite with thresholds, and the build fails on breach, no exceptions and no vibes. Online monitoring, run against production samples: catches drift that no frozen suite can, because the world changed under you. Teams usually build the first, sometimes the second, rarely the third, and the third is where the surprises live.
The gate is the point
A measurement that cannot fail a build is a dashboard, and dashboards get ignored. The discipline that changes behavior is the gate: defined thresholds, zero tolerance on critical scenarios, and a deliberately degraded prompt kept in a drawer to prove the gate actually catches things. The first time a well-meaning one-line prompt edit fails CI because two adversarial cases regressed, the whole team's relationship with the word better changes permanently.
The takeaway
Use metric words honestly, pass rate, precision, faithfulness, never a bare accuracy, and let no claim outrun its measurement. The eval harness is not bureaucracy around the AI. It is the part of the system that lets everything else move fast, because it is the part that knows when something broke.