Citation Check Against the Source Text

{
  "model": "jev-latest",
  "state": {
    "claim": "{{claim}}",
    "cited_section": "{{cited-section}}"
  },
  "questions": {
    "relation": {
      "type": "choice",
      "instructions": "How does the cited section relate to the claim?",
      "criteria": {
        "supports": "The section states the claim or directly implies that it is true",
        "contradicts": "The section states the opposite of the claim or implies it is false",
        "says_nothing": "The section does not address what the claim asserts, either way"
      }
    },
    "overstated": {
      "type": "noul",
      "instructions": "The claim is stronger, broader, or more certain than what the section actually says"
    }
  }
}

How to run it: before calling the model, check in code that any quoted words in the claim appear verbatim in the section after whitespace normalisation. A missing quote is a fabricated citation and needs no model call. Then map relation.choice to a verdict: supports becomes verified, contradicts becomes contradicted, says_nothing becomes unsupported. Accept a verdict automatically only when relation.confidence is 0.8 or above, and flag overstated above 0.6 as a separate warning even when the citation is verified. Start the threshold high and lower it as you see how the model does on your own documents.