Jev Gallery

JEV GUIDE

Jev citation verification: evidence extraction and claim support

An accessible citation does not necessarily support its surrounding claim. Jev can assess a claim against a passage, while retrieval, reference matching and the final research judgment require their own checks.

Citation review involves three distinct questions

Editorial recommendation: separate reference identity, the existence of an evidence passage in the retrieved text, and whether that passage supports the specific claim. Failure of either earlier check should prevent a “supported” result.

For example, improvement in one group under a particular condition does not automatically support effectiveness for everyone. Preserve population, conditions and comparison scope rather than equating shared subject matter with agreement. This is a synthetic example.

How do existing cases combine models and code?

Paper Trellis Citation Verifier pairs citing sentences with retrieved papers. Claude can propose evidence, code checks quotation existence and numbers, Jev assesses support, contradiction or silence, and the user records the final verdict. Its model thresholds have not been validated on a labeled set.

Source: Paper Trellis Citation Verifier · pinned source ↗

Keep abstracts, full text and extraction failures distinct

These are suggested editorial result states, designed to make the inspected evidence explicit.

Keep abstracts, full text and extraction failures distinct
Available materialAppropriate result
Abstract only, claim not mentionedFull text needed; current evidence insufficient
Full text with a matching passageShow passage, location and relation for human review
Reference matching or extraction failedUnable to verify; preserve the failure reason

What would a meaningful evaluation look like?

Jev Reviewer offers another route: selecting and checking evidence in numbered text. Browser-side file parsing does not mean the text used for inference stays out of model requests.

Source: Jev Reviewer · pinned source ↗

Preserve a reviewable evidence record

Editorial recommendation: retain the paper revision, inspected scope, passage location, model configuration and human verdict for each claim. Label examples of correct citations, overstated scope, contradictions and unavailable full text, then assess false support and false rejection separately.

Another model agreeing is not a sufficient ground-truth label. Researchers still need to inspect the original evidence. We have not measured these projects’ citation-review accuracy.

Explore these projects next

Continue reading