Aglet

Triage Grounding in Agent Answers

A fluent answer may contain details that the supplied context never supports. Triage grounding by defining the evidence boundary before reading the answer, then separate factual support from completeness or style. Include no-answer and conflicting-context cases so unsupported confidence is visible.

Establish what is happening

  1. Define supported claims

    State what the answer may infer from the supplied context, which claims need direct support, and what counts as an omission or contradiction. Record whether outside knowledge is allowed. Keep request fulfillment separate from grounding.

  2. Map context variation

    Include short and long documents, relevant distractors, conflicting sources, incomplete evidence, and unanswerable questions. Record document version, retrieval set, prompt, and answer format. Mark cases where a concise answer can omit detail without becoming ungrounded.

  3. Bound the review

    Specify claim extraction, support labels, evaluator view, abstention treatment, and decision use. Identify context or domain limits. Do not let a high fluency or relevance impression substitute for claim-level evidence.

What to carry forward

The triage output is a grounding contract with context, claim, support, omission, contradiction, and abstention rules. Stop when reviewers know what evidence is allowed. Keep a result provisional when context provenance or claim boundaries are unclear for later 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