Sample content for editorial review. Examples are illustrative.
There are several different citation failures
A reference can point to a work that does not exist. It can identify a real work with incorrect metadata. It can also point to a real, correctly described work that does not support the statement beside it. Treating all three as one problem makes evaluation and repair harder. I would design checks for these failure types separately.
Use a verifiable source record
Start with reference records obtained through the sources your workflow is authorized to use. Preserve the identifiers and metadata returned by those sources rather than asking the model to reconstruct them from memory. When an identifier or title cannot be resolved, mark the reference for review. Similar wording is a clue for investigation, not sufficient evidence that two records describe the same work.
Check the metadata without rewriting history
Compare the cited title, author information, publication venue, and date with the available record. Explain which fields disagree and preserve the original generated reference for debugging. Do not silently replace a questionable citation with the nearest plausible match. That may produce a convincing bibliography while disconnecting the paragraph from the evidence it was intended to describe.
Inspect support at the claim level
A successful metadata check does not establish that the cited work supports a claim. Link important statements to relevant source passages when access and permissions allow it. Review whether the passage contains the result, qualification, or limitation being described. Keep missing full text and uncertain support visible to the editor rather than translating them into a clean pass or fail without explanation.
Build a useful editorial queue
Group flagged references by the kind of review they need. An unresolved identifier, inconsistent metadata, and unsupported interpretation should lead to different next steps. Track which flags editors confirm and which create unnecessary work. This illustrative workflow aims to make checking easier and preserve the evidence behind corrections. It should help an editor decide what needs attention rather than imply that a valid-looking reference guarantees a valid scientific claim.
