Aglet

Investigate Unsupported Claims in an Agent Answer

Grounding investigation follows each important claim back to the evidence available at answer time. Split compound statements, compare source passages, and record missing support. Keep a wrong retrieval set distinct from an answer that ignored the right context during review.

Build a useful investigation brief

  1. Extract the claim set

    Copy material claims from the answer and record the request, context version, retrieved passages, and any abstention language. Mark qualifiers, dates, quantities, and causal wording that require stronger support than a general summary.

  2. Match claim to evidence

    For each claim, identify direct support, valid inference, contradiction, omission, or no support. Compare passage wording and scope. Check whether the answer fulfilled the request while exceeding the context boundary or leaving a required detail out.

  3. Test the evidence path

    Run a matched case with complete, conflicting, or missing context. Compare answer behavior and retrieval. State whether the failure is unsupported generation, missed evidence, context conflict, or an evaluator rule that cannot distinguish the cases.

What to carry forward

The investigation brief should contain claim map, source passages, support labels, context version, matched case, and competing explanations. End with a localized grounding failure or evidence gap. Do not call a claim grounded because it sounds plausible outside the supplied context.

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