AI evals that actually help (and the ones that don't)
Why most eval dashboards are theater, and what to measure instead.
Most eval dashboards measure things that look scientific and change nothing. BLEU on a customer support chatbot. Cosine similarity between a generated answer and a reference that no human ever read. Pretty numbers, zero signal.
The evals that move the needle are small, opinionated and tied to a real failure mode. 'Does the agent refuse when asked about refunds outside policy?' is a better question than 'what's the average answer quality?'. Twenty hand-labeled examples beat two thousand auto-generated ones every time.
Build the eval set the day you start the prompt, not the day before launch. Every regression you catch in dev costs a hundred times less than the one you catch in prod.