Define a Lightweight Evaluation Rubric You Can Apply to Any Run

Two support workflow runs can both read fluently and still behave very differently in operations. In one run, the triage JSON parses, category is billing, priority is high, next_action is escalate, and needs_human is true, so the router sends the case to the right queue. In the other run, the drafted reply sounds helpful, but the triage JSON contains priority set to urgent and a sentence cites a policy detail that does not appear in the retrieved snippets, so the run either fails validation late or, worse, passes through with a risky claim. Where the test table checks specific inputs against specific expected outputs, this lesson gives a rubric for judging any ...