How would you evaluate grounding, unsupported claims, and abstention?

The expected behavior depends on risk and source authority: a creative assistant may tolerate broad synthesis, while a policy or medical workflow needs narrower support and explicit uncertainty. I track unsupported-claim severity, evidence coverage, citation validity, correct abstention, false refusal, and conflict handling by slice. Production sampling includes the exact evidence available at response time so later source changes do not rewrite history. Severe unsupported claims can trigger a rollback or kill switch even if the aggregate answer score remains high.