Lessons learned: what 18 months of eval data actually predicts

The first time LLM evals in CI fell over in front of a customer, I had a dashboard that told me nothing. That is the story below.

The short version

blocking regressions in CI saved roughly nine rollbacks a month

What we tried

Move	What happened
Bigger context window	Cost up 3x, accuracy flat
Prompt rewrite on its own	+4% on the golden set, then regressed two weeks later
Swapping promptfoo for a boring alternative	Same quality, fewer moving parts

What broke

a 200-row golden set is enough; chasing 5,000 examples delays the actual fix

Behind almost every incident was a default we inherited and never questioned. A token limit too small. A timeout copied from an HTTP tutorial. A retry loop with no ceiling.

By the numbers

• p50 looked fine the entire time; p99 is where users actually lived
• 60% of "the model is wrong" tickets were retrieval returning the wrong document
• Rolling back a prompt was faster than fixing one — every single time

What I would do differently

1. Instrument first: inputs, outputs, latency, token counts, on day one
2. Build a 200-example golden set before the second feature, not after the fifth
3. Make every external call idempotent and deadline-aware
4. Read my own traces for an hour a week, on the calendar

The uncomfortable lesson

Most of what we shipped was ordinary backend engineering wearing an AI hat. None of this was clever. It was just the work nobody wanted to do first.