Done Is a Function You Write

The new unit test, for AI that thinks.

Your large language model crushed the benchmark. Then it shipped, and the support tickets started. Sound familiar? A nondeterministic system can be brilliant in the demo and wrong at scale, and you cannot regression-test it the way you test real code. The number you were steering by was measuring someone else's problem.

Done Is a Function You Write is the field manual for LLM evaluation, the discipline that quietly became the center of AI engineering. Its core move is a reframe: "done" is not a vibe or a leaderboard rank, it is a function you write. Once you can write that check, you can delegate to the machine everything that passes it.

Eval-driven development, from your first failing check to a suite that guards every deploy:

• Done as a function replaces the vibe check and the leaderboard rank with an executable definition of correct, treating evals as test-driven development for generative AI.
• Trace-first AI testing starts where the real failures live, your own production traces, so the suite measures your problem instead of a public benchmark's.
• LLM-as-judge alignment calibrates a machine grader against human labels and teaches you to spot and debug judge bias before it decides a release.
• Contamination defense shows why a verified score can drop thirty-five points overnight, and how to keep AI model evaluation honest when benchmarks leak into training data.
• Capability versus reliability separates what a model can do once from what it does every time, the distinction that decides what is safe to ship.
• The regression gate assembles a suite that blocks silent regressions on every deploy, the new unit test wired into your pipeline.

This is hands-on AI engineering, not theory. Read it and you will write the checks that decide what a thinking machine is allowed to do, delegate exactly as much as those checks prove safe, and ship AI you can stand behind. That is the scarce skill no model upgrade erases.

For engineers, data scientists, and applied-AI teams building with large language models and agentic AI, past "can the model do it" and stuck on "can I trust it enough to ship." Part of the Build Agents You Can Trust series, in The Verifier's Library.