Two-method grading layer

Implement exact or fuzzy matching for factual questions, and LLM-as-judge against an explicit written rubric (correctness, citation accuracy, refusal-appropriateness) for the rest. Output per-question pass/fail plus a category label.

Two methods because one does not fit. Exact match is worthless for synthesis and perfect for lookups; a judge is the reverse. Choosing per category - and writing the rubric down - is the actual skill.