Use six questions for every response

1. Supported

Can a reviewer trace every employer fact to reviewed information in the exact release?

2. Relevant

Does the response address the question asked rather than reciting nearby content?

3. Within scope

Does it respect the configured product, policy, safety and employment-decision boundaries?

4. Honest about uncertainty

Does it acknowledge absent, stale, ambiguous or conflicting information without filling the gap?

5. Clear about the next step

Where the reviewed material supplies one, does it identify an official source, process or support route?

6. Accessible and readable

Is it understandable in context, usable with assistive technology and available through a suitable alternative route?

Record evidence, not just a score

For each agreed question, keep the input, response, source evidence, review result and a concise reason. Record the exact content and configuration revision. A percentage without that context hides whether one serious unsupported claim sits behind many easy passes.

Include difficult questions deliberately

A representative set includes routine questions, paraphrases, ambiguity, edge cases, missing information, sensitive topics and requests outside the intended capability. Test the handoff and refusal behaviour as carefully as the ordinary answers.

Separate answer evaluation from candidate evaluation

This rubric evaluates the quality and boundary behaviour of the service. It must not become a way to score a candidate, infer suitability or make a hiring decision.

Keep live signals in their place

Manual evaluation can inspect exact agreed questions and answers. Where verified aggregate reporting is enabled, fixed-taxonomy themes and outcomes can suggest where to investigate. They do not reveal exact live questions or prove why a content gap occurred.

Further reading