Claim grounding in retrieval

A retrieval grounded assistant answers from passages fetched per query. In the settings it is sold into, a claim the evidence does not carry is a defect even when the claim happens to be true: a support agent quoting policy, an analyst citing a filing, a clinician summarizing a chart. Your job is a per claim verdict over the model's own output, supported or not.

This is not the attribution question. Nobody is asking which passages produced the sentence. The claim may have come entirely from the model's parameters, and what you have to decide is whether the retrieved evidence carries it.

The confound is that any judge you can afford already knows the world. It marks a true claim supported because the claim is true. It marks a claim that contradicts the passage unsupported because the claim is false. Both are the wrong reason, and the second case is the commercially important one: when a passage says something incorrect and the answer repeats it, that claim is supported, and your detector has to say so.

Constraints: no ground truth support labels, no retraining, and a per claim budget too small for a deliberation loop on everything. Do not assume one claim maps to one passage. Support may require two passages combined, an arithmetic step, or a date comparison. The decomposition of a sentence into claims is yours to specify, and it is not free, because a decomposer drops the hedges and quantifiers that carry the truth conditions.

Hand in a definition of supported precise enough that two engineers labelling the same pair agree, including where unstated but reasonable inference sits. Hand in the decomposition rule and what it does to a hedged or conditional sentence. Hand in the detector and its per claim cost. Hand in a separation experiment containing claims that are true and unsupported and claims that are false and supported, showing your verdicts follow the passage rather than the world or the fluency of the sentence. Name the claim family you systematically get wrong.

Ties break toward the crisper definition, since a detector is only as gradeable as the property it claims to detect.