What genuinely connects Dahl et al. (2024)'s 58%-legal-hallucination finding and Chen, Liang & Wang (2026)'s bare citation of Naser?
Checked 10-inbox/raw/ for a sibling capture mentioning both
claim-dahl-2024-llms-hallucinate-legal-queries-at-least-58-percent and
claim-chen-liang-wang-2026-cites-naser-as-bare-reference-not-empirical-test
together: none found (two files matched one slug each, none matched both).
This is a fresh investigation of the pair named in the topic.
Re-measured the pairing directly with vault_pair rather than taking the
topic brief's cosine on trust: cosine 0.882 (document-document basis),
unlinked, rank 2 among the Dahl note's own nearest neighbours (at the
99.9th percentile of its cosine distribution) and rank 11 among the
Chen-Liang-Wang (CLW) note's. The vault already holds a structurally
adjacent finding at almost the same cosine:
Ansari × CLW, cosine 0.888, which resolved as a clean embedding false
friend — Ansari's paper is not in CLW's reference list at all, and the
two describe opposite kinds of citation event. Given that precedent, the
working hypothesis entering this investigation was that Dahl × CLW would
resolve the same way. It does not. A direct read of CLW's own reference
list, archived fresh this session, finds Dahl et al. (2024) cited by full
bibliographic entry. This is a real citation link, not a shared-vocabulary
artifact — the rarer case in this neighbourhood.
Claim: Chen, Liang & Wang (2026) genuinely cite Dahl et al. (2024) in their own reference list — the bridge between the two notes is a real citation, not an embedding artifact
A direct archive of CLW's Springer article page (doi.org/10.1007/s11192-026-05727-6, same sha as already on file for the existing claim-note, confirming no drift) shows the paper's reference list contains a full entry for Dahl et al.'s "Large Legal Fictions," reproduced exactly as printed: author names, year, title, venue, volume/issue/pages, and DOI. This settles the question the topic poses at the most basic level — the two notes' underlying documents are not independent; one cites the other. Unlike the Ansari pairing, where an end-to-end read of the same 40-entry reference list found no entry at all for Ansari, Beltran et al., or "Contamination Inheritance," Dahl et al. is present, correctly and completely cited.
- source_url: "https://doi.org/10.1007/s11192-026-05727-6"
- source_author: "Chung-Kuang Chen, Shu-Hao Liang, Teng-Shih Wang"
- source_date: "2026-07-22"
- source_title: "Beyond existence checks: a cross-database benchmark of bibliographic metadata consistency across academic domains and LLM-generated citations"
- source_venue: "Scientometrics, vol. 131, pp. 5965–5984 (2026), Springer"
- source_quote: "Dahl, M., Magesh, V., Suzgun, M., & Ho, D. E. (2024). Large legal fictions: Profiling legal hallucinations in large language models. Journal of Legal Analysis,16(1), 64–93. https://doi.org/10.1093/jla/laae003"
- source_tier: 1
- source_sha: "fc3429652277732f1cda94fe4d8a47c1c753f00a440eae3bf3bd2d2eb5ba710f"
Claim: CLW does not test or replicate Dahl's 58% hallucination-rate finding — it cites Dahl in the same bare-reference shape it already uses for Naser, extending that pattern rather than breaking it
CLW's own abstract states its measured target plainly: a cross-database benchmark of bibliographic metadata consistency, not of hallucination rate. Its corpus is "1,246 citations ... queried against three scholarly databases (CrossRef, OpenAlex, and Semantic Scholar) across seven metadata fields and five academic domains," and its headline result is that "66.7% of all citations exhibit at least one metadata disagreement between databases." Nothing in the abstract, the data-availability note, or the reference list indicates CLW re-runs, replicates, or empirically tests Dahl's reported "at least 58%" legal-hallucination figure — Dahl's citation sits in the same kind of related-work company as Naser's: domain- or model-specific audits of LLM citation/reference failure, cited to motivate the problem CLW's own (differently-shaped) study addresses. The reference list groups Dahl alongside several other domain-specific hallucinated-citation studies cited the same way (medical: Bhattacharyya et al. 2023, Chelli et al. 2024, Gravel et al. 2023; mental health: Linardon et al. 2025) — Dahl is one bare reference among a cluster of them, not a singled-out test target.
This is the same shape the vault already recorded for CLW's citation of Naser: "Nothing in its visible abstract, data-availability note, or reference list indicates it tests or replicates Naser's reported 11.4%–56.8% hallucination-rate finding." The genuine connection the embedding caught (cosine 0.882) is that both notes describe adjacent entries in the same reference list — CLW is a hub paper that cites multiple independent LLM-citation-hallucination measurement studies (Dahl's legal-domain rate, Naser's cross-model rate, among others) as background, without empirically testing any one of their specific figures. [unverified-mechanism — needs primary]: Springer serves only an abstract/reference-list preview to this vault's tooling ("This is a preview of subscription content, log in via an institution to check access"), so the exact in-text sentence and section where Dahl is cited (introduction vs. elsewhere) could not be directly confirmed — the "bare reference, not test" characterization rests on comparing CLW's stated empirical target against Dahl's, the same evidentiary standard the vault's existing Naser note already used for the same paper.
- source_url: "https://doi.org/10.1007/s11192-026-05727-6"
- source_author: "Chung-Kuang Chen, Shu-Hao Liang, Teng-Shih Wang"
- source_date: "2026-07-22"
- source_title: "Beyond existence checks: a cross-database benchmark of bibliographic metadata consistency across academic domains and LLM-generated citations"
- source_venue: "Scientometrics, vol. 131, pp. 5965–5984 (2026), Springer"
- source_quote: "A corpus of 1,246 citations—comprising 491 AI-generated references from five large language models (Claude Opus 4.6, GPT-4.1-nano, Gemini 3 Flash, Llama 3.3 70B, and Llama 3.1 8B) and 755 database-sampled references—was systematically queried against three scholarly databases (CrossRef, OpenAlex, and Semantic Scholar) across seven metadata fields and five academic domains."
- source_quote_2: "Results reveal that 66.7% of all citations exhibit at least one metadata disagreement between databases, with author names showing the lowest inter-database agreement (29–39%) and DOIs the highest (100%)."
- source_tier: 1
- source_sha: "fc3429652277732f1cda94fe4d8a47c1c753f00a440eae3bf3bd2d2eb5ba710f"
Claim: This resolves differently from the vault's nearest precedent at almost the same cosine — the Ansari × CLW pairing is a false friend (no citation exists), while Dahl × CLW is a confirmed bridge (the citation exists and is verified)
The two pairings sit a hair apart in cosine (0.888 for Ansari × CLW, 0.882 for Dahl × CLW) and both involve the same CLW note as one leg, which is why they are worth distinguishing explicitly rather than assuming the first result generalizes. The Ansari pairing resolved as a false friend on two independent grounds: no shared reference (Ansari absent from CLW's full reference list) and a chronology that forecloses mutual citation. Neither ground applies here: Dahl is in CLW's reference list, correctly cited, and chronologically CLW (published 22 July 2026) postdates Dahl (2024) by two years, so the citation is unremarkable rather than impossible. The distinguishing fact is not cosine — the two scores are nearly identical — but a direct read of the cited document, which is the only test that can tell a true bridge from an embedding false friend in this cluster, as the vault's own prior write-ups of this cluster (observation-ansari-naser-2026-cosine-080-pairing-is-analytical-contrast, observation-ansari-chen-liang-wang-contamination-inheritance-cosine-089-pairing-is-embedding-false-friend) already establish as the operational standard.
claim-dahl-2024-llms-hallucinate-legal-queries-at-least-58-percent and claim-chen-liang-wang-2026-cites-naser-as-bare-reference-not-empirical-test should be wikilinked on this basis: CLW's document is a confirmed, direct citer of Dahl's document, even though — per the second claim above — it cites without testing.
- source_url: "https://doi.org/10.1007/s11192-026-05727-6"
- source_title: "Beyond existence checks: a cross-database benchmark of bibliographic metadata consistency across academic domains and LLM-generated citations"
- source_tier: 1
- source_sha: "fc3429652277732f1cda94fe4d8a47c1c753f00a440eae3bf3bd2d2eb5ba710f"
- source_quote: "vault_pair tool reading (direct instrument measurement, not a document): cosine=0.882, rank_of_b_among_a_neighbours=2, rank_of_a_among_b_neighbours=11, percentile_within_a=99.9, linked=false"
- source_tier_quote: 1
- source_url_2: "vault:30-notes/observation-ansari-chen-liang-wang-contamination-inheritance-cosine-089-pairing-is-embedding-false-friend.md"
- source_title_2: "The Ansari Contamination-Inheritance × Chen, Liang & Wang (Naser-citing) cosine-0.888 pairing is an embedding false friend"
- source_tier_2: 1
Further leads
- CLW's abstract names only three of its "five academic domains" with figures (biomedicine 75%, social science 72%, CS/ML 61%); the other two are not visible in the paywalled preview. Worth a later check of whether "law" is one of the five — that would tie Dahl's domain to CLW's own sampling frame, not just its background citations.
- CLW's 40-entry reference list is a hub for the wider LLM-citation-hallucination literature beyond Dahl and Naser: Bhattacharyya et al. 2023, Chelli et al. 2024, and Gravel et al. 2023 (medical-domain hallucinated citations), Linardon et al. 2025 (mental-health domain), Xu et al.'s GhostCite (arXiv:2602.06718), Walters & Wilder 2023, Agrawal et al. 2024, and Sakai et al. 2026 (arXiv:2601.18724) all appear — worth a systematic sweep for other vault claim-notes that might form the same "bare reference, not test" pairing against CLW.
- The exact in-text sentence/section where CLW cites Dahl (e.g. introduction's motivating survey of domain-specific hallucination work) is unconfirmed — Springer's article page serves only an abstract-and-references preview to this vault's tooling; a full-text route (institutional access, author-hosted copy) was not found this session.
Entity candidates
- Matthew Dahl, Varun Magesh, Mirac Suzgun, Daniel E. Ho — persons — the Yale/Stanford RegLab group behind "Large Legal Fictions"; no entity page found despite being the older, foundational finding both this pairing and the vault's broader citation-hallucination cluster measure newer papers against (flagged first, per the standing blind-spot note, not last — the vault has already built an entity page for Naser, the newer citer, but not for Dahl, the earlier measured-against figure).
- Chung-Kuang Chen, Shu-Hao Liang, Teng-Shih Wang — persons — authors of the CLW paper itself (National Taiwan University of Science and Technology; Providence University); no entity page found despite the paper now anchoring two confirmed-vs-false-friend bridge investigations in this cluster.
- "Bare bibliographic reference, not empirical test" — concept — a recurring pattern this cluster keeps finding (CLW→Naser, CLW→Dahl, and per the vault's existing notes also CPL→NeuNeu and Chereches→Ansari); may be worth its own concept note if a fourth instance lands, rather than re-describing the pattern each time from scratch.