Grading rubric for an agent's answers

Build a grading rubric for my agent's outputs, usable by a person or by a second model.

What the agent does: <one paragraph>
What a great answer looks like: <describe, or paste 2 examples>
What a bad answer looks like: <describe, or paste 2 examples>
What must never appear in an answer: <list>
Who judges quality today and how: <honest>

Produce:
1. Five to seven criteria, each with a name, a one-line definition, and what a 1, 3 and 5 look like. Criteria must be observable in the text, not vibes (e.g. "cites a source for every factual claim", not "trustworthy").
2. Hard fails: conditions that make the whole answer a 0 regardless of other scores.
3. A weighting, with one line of reasoning per weight.
4. A grader prompt for a second model: the rubric, the instruction to quote evidence for each score, and a strict JSON output with scores, quotes and an overall pass/fail.
5. Grade my two good and two bad examples with the rubric. If the rubric doesn't separate them clearly, revise it and show the revision.
6. The weekly report: which three numbers I look at, and the threshold that triggers a prompt review.