Evals or it didn't happen: how we ship AI features without burning trust
An evaluation rubric isn't a benchmark you publish — it's a discipline you keep. Here's the eval flow we now run on every production AI surface, and the three failure modes it catches before customers do.
Most teams treat evals as a one-shot benchmark you run before launch. We treat them as a maintenance discipline — the same way we treat tests for any production system.
The flow
For every AI surface we operate, three eval suites run continuously: a regression suite (built from real, anonymised support tickets), a rubric suite (judgment-graded by the buyer's domain expert), and a drift suite (a fixed seed-set rerun against the current model+prompt combination weekly).
Three failure modes it catches
- Silent regression: a prompt change that lifts the rubric and tanks the regression suite. We see this on every fourth merge.
- Distribution shift: input patterns the original eval set never saw. The drift suite is how we catch it before a customer reports a "weird answer".
- Reward hacking: the model getting better at the rubric and worse at the underlying job. Only the regression suite catches this.
What we don't do
We don't publish public benchmark numbers. They're not load-bearing for our customers and they're easy to game. We share the methodology, not the score.