The eval pipeline that finally caught our hallucinations.

We used to ship LLM features the way everyone does: prompt, demo, deploy, watch the dashboard, and hope. Then we stopped — because the dashboard only ever caught the loudest regressions. The quiet ones cost us a client.

What the eval pipeline actually looks like

Three layers, all running in CI before anything reaches production.

1. Golden cases — 50 to 200 hand-graded inputs with expected outputs. Run every PR.
2. Critic models — a second LLM judging answers on a fixed rubric. Cheaper than humans, faster than waiting.
3. Production traces — sampled real traffic, replayed against the new model. Catches drift the golden set never anticipated.

What it caught

On Renoir Labs' copilot, the eval pipeline flagged a 14% regression in answer fidelity the day we tried a 'smaller, cheaper' model. Without it, that would have shipped, and the analysts would have eventually stopped trusting the tool. With it, we rolled back inside the hour.