What this article gives you: How to design LLM quality evaluation around "binary checks + a named verdict" instead of numeric scores. Includes a judge prompt structure you can reuse by swapping in your own evaluation target.
Having an LLM evaluate deliverables is called LLM-as-judge. Once you introduce it, most people run into the same three walls:
Ask for a 5-point rubric score and the same input oscillates between 3 and 4 across runs
Gate pass/fail on a total-score threshold and one fatal flaw gets diluted by high scores on everything else — and slips through
Read the scores back later and nobody can explain why something got a 3.5
