Before you write good evals, you have to pick the right things to evaluate. Main take-aways:
- source metrics from failures you see in traces
- keep the set small and maintainable
- only keep a metric if you'd take action when it moves x.com/i/article/2089…