Check whether the outcome improved
Set abstention checks
Require answerability, evidence sufficiency, intent, response choice, unsupported-claim, clarification, partial-help, and next-action rules. Define acceptable qualification and refusal for independent reviewers. Mark unavailable context as an evidence limitation for each tested case.
Review boundary cases
Apply checks to complete, missing, conflicting, ambiguous, out-of-scope, and safely answerable requests. Inspect claims against context and the user next step. Record both a justified abstention and a case where refusal was unnecessary.
Approve response scope
State which tasks, evidence conditions, and response choices the result supports. Record reviewer, context version, owner, and retest trigger. Mark results inconclusive when answerability or user intent cannot be established.
What to carry forward
The verification record should show answerability, context, response choice, claim support, next action, and limits. Approve only the abstention behavior observed. Reopen it when evidence sources, request types, prompts, or response rules change. Document whether the user could continue with the supplied next action after this review.
Technical background: Google DeepMind evaluation research.
Keep the decision with the work.
Use a Work Item in Aglet to record the problem, the evidence you have, and the next decision. Add an owner and priority, then keep updates in the discussion so the next person can follow the reasoning.
Create an account See the product workflow