How would you start evaluating an existing feature with no evals?

An answer that works

Read real traces first and write down how it fails, so the eval measures actual problems. Then take twenty real inputs, including ones that have failed, write down what a good output looks like for each, and score the current system to get a baseline. That number is what every future change is measured against.