A clear rubric turns vague judgements of whether an answer seems good into repeatable, decision-relevant review.
Define the task before selecting metrics
A chatbot drafting a friendly email and a system summarizing medication instructions require different grading criteria. Start by writing what the output must achieve and which errors matter. Separate factual accuracy, completeness, citation support, instruction compliance, harmlessness, accessible formatting and user usefulness. Do not force all dimensions into one number if a severe error should block deployment regardless of style.
Write observable scale anchors
Use anchors that reviewers can apply consistently: for citation support, ‘fully supported’ means each decision-bearing statement is traceable to an exact passage; ‘partially supported’ means nonessential statements lack evidence; ‘unsupported’ means material claims do not follow from the sources. Supply examples of borderline cases. Vague labels such as outstanding or poor leave too much room for individual taste.
Control reviewer disagreement
Recruit reviewers who understand the task and the language. Blind them to vendor names where practical. Give them the same instructions and independently score a sample before comparing results. Examine disagreements to identify ambiguous instructions or flawed ground truth. Report both the rate of agreement and substantive disagreement patterns instead of hiding them behind a mean.
Protect independence
If a reviewer has access to the model’s preferred answer, anchoring can bias judgement. Randomize the order of outputs for comparisons and remove vendor branding where feasible. Keep answer keys separate from evaluators until the scoring round closes. Store the rubric version, case list and adjudication decisions so later changes can be explained.
Weight errors by consequence
A missing comma and a wrong emergency instruction should not cancel out as one error each. Introduce blocker criteria for unsafe guidance, unauthorized disclosure and fabricated legal or medical citations. For other categories, define a severity scale and specify who can approve exceptions. Reporting refusal and escalation frequency helps prevent a system from gaming accuracy by declining all difficult cases.
A sample mini-rubric
For an internal policy answer, a five-point review could assess: correct controlling document, accurate conclusion, quoted exception, correct effective date, and acknowledgement of missing data. Each point is binary with a short justification, while incorrect authorization advice is a separate blocker. Reviewers then provide a brief issue tag rather than rewriting the whole answer during scoring.
Operational limitations
Human review is itself inconsistent, time-consuming and subject to bias. Small numbers of cases cannot establish reliable group differences. Periodically audit review quality and involve affected users, especially when accessibility or language coverage could change how an apparently correct answer is experienced.
Sources & Further Reading
- NIST AI RMF Core: Govern Map Measure Manage — NIST.
- Holistic Evaluation of Language Models — Stanford CRFM.
- NIST AI Risk Management Framework 1.0 — NIST.
Primary Scope and applicability vary by task and jurisdiction. This article is educational guidance, not legal, medical or financial advice.