A practical rubric for checking documentation answers

Record the answer before judging it

Save the question, response, cited URL, source version and configuration date in an approved review record. Remove personal information that is not needed for the test. Without the actual output, a rating such as good or bad is difficult to reproduce. Keep the expected answer points beside the response so the reviewer can identify precisely what is missing, contradicted or invented.

Use separate review dimensions

Check whether essential facts are correct, whether each important claim is supported by the cited material, whether the version is appropriate and whether the response stays inside its remit. Then assess clarity and the next useful action. These dimensions should not collapse into a single attractive score. An answer can be well written and materially unsafe to follow, or accurate but too vague to help.

Apply a fictional example

Imagine a source that permits CSV export only for workspace administrators. An answer that describes the correct menu but says every user can export fails the permission fact. An answer that omits the restriction is incomplete even if its citation points to the correct page. The example is deliberately small: it shows how to review a claim, not how any named chatbot performed in our testing.

Keep the decision and the reason

Classify each case as acceptable, needs repair, unsupported or unresolved, with a short explanation and the affected source passage. Use a second reviewer for disputed or consequential cases. Do not treat agreement between two models as independent factual verification. The worksheet is a structured aid for human source review; it cannot certify correctness automatically or replace product-specific expertise.

• For each held claim, write the exact missing or conflicting passage and the smallest correction that would change the judgment.