---
id: "20261116-0206-bipartite-two-things-this"
title: "Bipartite hop: Beltran et al. v1's traced fabricated citation and Naser (2026)'s no-Frankfurt citation-fabrication audit — a real mechanism bridge through Ansari, not a false friend"
type: "capture"
status: "promoted"
origin: "batch"
promoted_to: ["30-notes/observation-ansari-naser-2026-cosine-080-pairing-is-analytical-contrast.md","30-notes/claim-ansari-2026-zhu-neuralrgbd-is-her-tf-pac-boundary-case.md","40-entities/entity-azinovic-et-al-2022.md (new hub)","40-entities/entity-walters-wilder-2023.md (new watching stub)","40-entities/entity-mz-naser.md (dated update)","40-entities/entity-samar-ansari.md (dated update)","50-questions/question-corroborate-ansari-naser-2026-single-source-cap.md (routed, not promoted as a claim)"]
not_promoted: ["Claim 3 (Naser's audit is cited-from and citing-to nobody in this cluster — mutual non-citation with Ansari, widening the Frankfurt gap): the underlying fact was directly verified this session (both reference lists read end to end) but promoting it as a claim-note would push BOTH Ansari's (arXiv:2602.05930) and Naser's (arXiv:2603.03299) claim-note counts to 4, past the single-source concentration cap of 3 — each already reached exactly 3 via the two notes that were promoted. Routed instead to question-corroborate-ansari-naser-2026-single-source-cap.md per sources.md's cap mechanism.","Further lead: Naser's related-work citation of Walters & Wilder (2023) as the founding empirical study of LLM citation fabrication — not a claim of this capture's own, just a pointer; seeded a `status: watching` entity stub (entity-walters-wilder-2023.md) rather than a claim-note, since the paper itself was not read directly this session.","Further lead: whether Naser's publicly-released gradient-boosted classifier would score the Zhu/NeuralRGB-D citation's own bibliographic string features as 'hallucinated' — not tested this session, no claim to promote.","Further lead: whether Ansari's §4.7 TF/PAC ambiguity rate is quantified anywhere beyond the single traced case — not checked this session, no claim to promote.","Further lead: Naser's within-family Jaccard convergence and the haiku-4.5-regressed-relative-to-haiku-3.5 finding — tangential to this pair, not investigated, no claim to promote.","Entity candidate GPTZero: declined a hub this round. It acts (its automated Hallucination Check tool is the dataset-generating detector behind Ansari's audit) but on inspection has only ONE substantive load-bearing role in the vault so far — its appearance in the tortured-phrases cluster (claim-cabanac-2021-tortured-phrases-named-before-chatgpt) is a one-sentence backward cross-reference to the same Ansari role, not a second independent cluster where GPTZero itself acts. Fails the org-hub bar's 'more than one independent cluster' test as currently written; worth revisiting if GPTZero surfaces acting independently somewhere else."]
writer_model: "claude-sonnet-5"
date_created: "2026-11-16T00:00:00.000Z"
provenance: "batch run, 2026-11-16"
derived_from: []
tags: ["citation-fabrication","contamination-inheritance","samar-ansari","mz-naser","bridge-pair","embedding-bridge","hallucination-taxonomy","beltran-nfl-ba"]
source_urls: ["https://arxiv.org/pdf/2603.03299","https://arxiv.org/pdf/2602.05930","vault:30-notes/claim-beltran-v1-2024-contains-verbatim-neuralrgbd-fabrication-ansari-traced.md"]
source_tiers: ["Tier 1 (primary preprint, M.Z. Naser, re-fetched and re-verified this session)","Tier 1 (primary preprint, Samar Ansari, re-fetched and re-verified this session)","Tier 1 (Seek's own prior vault claim-note, primary-grounded, capture-verified via extract_pdf against Beltran et al. v1's own PDF)"]
seek_code_commit: "d40ee39"
---


Checked `10-inbox/raw/` for a sibling capture mentioning both note slugs together first: none found. One related-but-distinct bridge capture exists ([[observation-david-brachet-ansari-contamination-inheritance-cosine-086-pairing-is-hedged-bridge|2026-09-20's recovered capture]]), pairing [[claim-ansari-2026-contamination-inheritance-citation-error-propagates-across-models]] against a different seed note (David & Brachet), and ruling that pairing a hedged bridge. This is a fresh investigation of the pair actually named in this topic — [[claim-beltran-v1-2024-contains-verbatim-neuralrgbd-fabrication-ansari-traced]] and [[claim-naser-2026-citation-fabrication-audit-no-frankfurt-bullshitting-citation]] — not a duplicate of it, though it reuses the same Ansari primary document.

`mcp__seek__vault_bridge` on the pairing returned the two seed notes as each other's top-2 closest neighbors (cosine 0.804 and 0.798) but `bridge_candidate: false` — the frontier-density test the tool applies didn't flag it, despite the two notes visibly being each other's nearest neighbors. The notes share no author, no document, and (per the task's own framing) no vocabulary. Starting from the less familiar side — Naser's audit — and re-fetching both primaries directly (Naser: arXiv:2603.03299, sha256 `651f72d7...`, tls verified; Ansari: arXiv:2602.05930, sha256 `66d7e5cf...`, tls verified, matching the sha already on file in the vault with no drift) turned up a real, quantified, mechanism-level connection that the cosine score's vocabulary-matching couldn't see: the two papers independently built structurally parallel taxonomies of the same phenomenon, one day apart, and reached **opposite** headline conclusions.

## Claim: Ansari's and Naser's independently-built hallucination taxonomies, posted one day apart with no cross-citation, reach opposite conclusions on whether LLM citation fabrication is mostly invention or mostly corruption of something real

Ansari's paper (the source for the Beltran-v1 trace) analyzes 100 hallucinated NeurIPS-2025 citations by failure mode and concludes: "LLMs default to invention, not corruption... two-thirds of hallucinations involve wholesale invention with zero overlap with reality" — Total Fabrication is her largest category at 66%, versus Partial Attribute Corruption (blended real/fake elements) at 27%. Naser's paper, analyzing 69,557 citations generated by 10 different LLMs, scores each hallucinated citation's resemblance to a real publication and concludes the opposite: "This distribution infers that the dominant hallucination mechanism is not invention but distortion — models retrieve partial bibliographic information and imperfectly reconstruct it to produce outputs that closely resemble but do not exactly match real publications." In Naser's typology, citations with "no detectable resemblance to any indexed publication" (her closest analogue to Total Fabrication) are only 6.0% of the hallucinated set; the remaining 94% show some structural resemblance to a real paper (58.3% "probable real," 31% "near miss," 4.7% "distant chimera").

Ansari's paper is dated February 6, 2026 (arXiv:2602.05930v1); Naser's is dated February 7, 2026 (arXiv:2603.03299v1) — one day apart. Neither paper's reference list cites the other: Naser's complete 20-entry reference list (already checked end-to-end for [[claim-naser-2026-citation-fabrication-audit-no-frankfurt-bullshitting-citation|the Frankfurt/bullshitting absence claim]]) contains no entry for Ansari, Beltran, GPTZero's NeurIPS audit, or "Contamination Inheritance"; Ansari's complete 14-entry reference list likewise contains no entry for Naser. This is the mechanism the 0.86 cosine score was pointing at without being able to name it: not shared vocabulary (their category names — Total Fabrication/Partial Attribute Corruption versus fabrication/near-miss/distant-chimera — share no words), but a shared empirical question, answered oppositely, by two papers that never saw each other.

- source_url: "https://arxiv.org/pdf/2602.05930"
- source_title: "Compound Deception in Elite Peer Review: A Failure Mode Taxonomy of 100 Fabricated Citations at NeurIPS 2025"
- source_author: "Samar Ansari"
- source_date: "2026-02-06"
- source_venue: "arXiv preprint 2602.05930v1 [cs.DL], School of Computing and Engineering Sciences, University of Chester"
- source_sha: "66d7e5cf32e40acd7322c06d8653f9eb7d0db9c64f0fdf958cc625cccf539919"
- source_quote: "LLMs default to invention, not corruption. One might hypothesize that language models trained on scholarly literature would generate citations by retrieving and corrupting real sources. Our data contradicts this: two-thirds of hallucinations involve wholesale invention with zero overlap with reality."
- source_tier: 1

- source_url: "https://arxiv.org/pdf/2603.03299"
- source_title: "How LLMs Cite and Why It Matters: A Cross-Model Audit of Reference Fabrication in AI-Assisted Academic Writing and Methods to Detect Phantom Citations"
- source_author: "M.Z. Naser"
- source_date: "2026-02-07"
- source_venue: "arXiv preprint 2603.03299v1, cs.CL, Clemson University"
- source_sha: "651f72d795862878fe13c06a3afa54f64f91d6c155470e1c09f899de04e1314d"
- source_quote: "This distribution infers that the dominant hallucination mechanism is not invention but distortion — models retrieve partial bibliographic information and imperfectly reconstruct it to produce outputs that closely resemble but do not exactly match real publications."
- source_quote_2: "Across all models, 58.3% of hallucinated citations fell in the probable real category, 31% were near misses, 4.7% were distant chimeras, and 6.0% were outright fabrications."
- source_tier: 1

## Claim: The exact fabricated citation anchoring the Beltran-v1 claim is Ansari's own dataset's test case for the TF/PAC boundary her paper's future-work section says needs resolving

The citation both seed notes trace — "Z. Zhu, T. Yu, X. Zhang, J. Li, Y. Zhang, and Y. Fu. Neuralrgb-d..." — is presented by Ansari in her paper's Total Fabrication section (§3.2) as a TF example, but by her own account in the same paragraph it is not a clean fit for that category's definition ("the authors, title, venue, and identifiers are all fabricated... zero overlap with any real publication"): the citation carries a real venue (CVPR) and a real year (2022) matching an actual 2022 paper on near-identical subject matter — [[claim-beltran-v1-2024-contains-verbatim-neuralrgbd-fabrication-ansari-traced|the real paper is Azinović, Martin-Brualla, Goldman, Nießner & Thies's "Neural RGB-D Surface Reconstruction," an unrelated author team]] from the fabricated six-name list. Ansari's own Future Work section names this exact tension: "What proportion of citations we classified as Total Fabrication are actually Contamination Inheritances from contaminated training sources?" — with the Zhu/NeuralRGB-D case as her dataset's only confirmed instance of the newly-coined Contamination Inheritance failure mode, discovered only because she traced this one citation back to Beltran et al.'s v1 preprint rather than relying on her five-category taxonomy's output-level coding alone.

This is a second, independent point of contact between the two seed notes' underlying papers: the single citation [[claim-beltran-v1-2024-contains-verbatim-neuralrgbd-fabrication-ansari-traced|the Beltran-v1 claim]] verified verbatim is also the specific boundary case that exposes the TF/PAC ambiguity Ansari flags in her own limitations section ("the boundaries between categories (particularly PAC vs. TF) can be ambiguous when citations blend multiple real and fake elements") — and that ambiguity is structurally the same question Naser's paper answers at scale with the opposite headline number (Claim 1).

- source_url: "https://arxiv.org/pdf/2602.05930"
- source_title: "Compound Deception in Elite Peer Review: A Failure Mode Taxonomy of 100 Fabricated Citations at NeurIPS 2025"
- source_author: "Samar Ansari"
- source_date: "2026-02-06"
- source_venue: "arXiv preprint 2602.05930v1 [cs.DL], University of Chester"
- source_sha: "66d7e5cf32e40acd7322c06d8653f9eb7d0db9c64f0fdf958cc625cccf539919"
- source_quote: "A third example involves apparent computer vision content: Z. Zhu, T. Yu, X. Zhang, J. Li, Y. Zhang, and Y. Fu. Neuralrgb-d: Neural representations for depth estimation and scene mapping. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2022. The title uses appropriate terminology for CVPR, yet no matching paper exists in the 2022 proceedings."
- source_quote_2: "What proportion of citations we classified as Total Fabrication are actually Contamination Inheritances from contaminated training sources?"
- source_tier: 1

> [!note] Seek's commentary: The task said "they share no vocabulary by construction... if the resemblance is real it is a MECHANISM, not a word," and this is about as clean a confirmation of that framing as the vault has produced. Nothing in Naser's paper uses Ansari's category names; nothing in Ansari's paper uses Naser's scoring vocabulary. But both papers are answering the identical question — when an LLM fabricates a citation, is it inventing from nothing or corrupting something real? — and they answer it with numbers that don't just differ, they invert (66% vs. 6% for "no overlap with reality"). That's not a coincidence the embedding could see, since cosine similarity runs on words, not on "these two authors computed the same statistic and got opposite values." Whether the gap is a real empirical disagreement (different datasets: 100 NeurIPS citations vs. 69,557 prompted-LLM citations, which is itself a confound neither paper controls for) or a definitional one (TF vs. "fabrication" aren't drawn at the same threshold) is a further question this capture doesn't settle — flagged below rather than forced. — Seek

## Claim: Despite the mechanism-level contact above, Naser's audit remains cited-from and citing-to nobody in this cluster — widening the gap the existing claim-note already found on the Frankfurt side

[[claim-naser-2026-citation-fabrication-audit-no-frankfurt-bullshitting-citation]] already established that Naser's 20-entry reference list contains no citation to Frankfurt, Hicks et al., or bullshitting/interpretability literature. This session's full re-read of that same reference list (end to end, re-verified this session against the primary PDF rather than the existing note alone) confirms it also contains no citation to Ansari's paper, to Beltran et al.'s NFL-BA preprint, to GPTZero's NeurIPS-2025 audit, or to the term "Contamination Inheritance" — and the reverse direction holds too: Ansari's complete 14-entry reference list contains no citation to Naser. The one-day posting gap (2026-02-06 to 2026-02-07) makes a citation in either direction implausible on timing grounds alone, independent of the reference-list check. The vault's citation-fabrication research corner is not just isolated from the Frankfurt/bullshitting philosophical literature, as the existing note shows — its own most directly comparable pair of 2026 papers, addressing the identical question with opposite answers, are mutually invisible to each other too.

- source_url: "https://arxiv.org/pdf/2603.03299"
- source_title: "How LLMs Cite and Why It Matters: A Cross-Model Audit of Reference Fabrication in AI-Assisted Academic Writing and Methods to Detect Phantom Citations"
- source_author: "M.Z. Naser"
- source_date: "2026-02-07"
- source_venue: "arXiv preprint 2603.03299v1, cs.CL, Clemson University"
- source_sha: "651f72d795862878fe13c06a3afa54f64f91d6c155470e1c09f899de04e1314d"
- source_quote: "The empirical study of LLM citation fabrication began with Walters and Wilder [3], who prompted GPT-3.5 and GPT-4 to generate short literature reviews on 42 multidisciplinary topics and manually verified the resulting 636 references."
- source_tier: 1

- source_url: "https://arxiv.org/pdf/2602.05930"
- source_title: "Compound Deception in Elite Peer Review: A Failure Mode Taxonomy of 100 Fabricated Citations at NeurIPS 2025"
- source_author: "Samar Ansari"
- source_date: "2026-02-06"
- source_venue: "arXiv preprint 2602.05930v1 [cs.DL], University of Chester"
- source_sha: "66d7e5cf32e40acd7322c06d8653f9eb7d0db9c64f0fdf958cc625cccf539919"
- source_quote: "Acknowledgments\nThe author thanks the GPTZero team for making their NeurIPS 2025 hallucination analysis publicly available."
- source_tier: 1

## Further leads

- Naser's related-work section names Walters & Wilder (2023, *Scientific Reports*) as the paper where "the empirical study of LLM citation fabrication began" — a founding-document candidate for a lineage note on this whole sub-literature, not followed up here.
- Naser's gradient-boosted classifier (AUC 0.876 CV / 0.834 leave-one-model-out) is trained on bibliographic string features including author count and "et al." presence; whether applying it to the Zhu/NeuralRGB-D citation's own string features would score it as "hallucinated" was not tested — the classifier and dataset are both publicly released (HuggingFace `mznaser/llm-citation-hallucination-audit`), so this is a checkable follow-up, not performed this session.
- Ansari's §4.7 Limitations explicitly flags "the boundaries between categories (particularly PAC vs. TF) can be ambiguous" as a general property of her coding scheme, independent of the Zhu/NeuralRGB-D case — whether this ambiguity rate is itself quantified anywhere in her paper (beyond the single traced case) was not checked.
- Naser's paper documents a within-family convergence pattern (Jaccard similarity between sibling models) and a "haiku-4.5 regressed relative to haiku-3.5" generational finding — tangential to this pair, not investigated as part of the bridge.

## Entity candidates

- Azinović, Martin-Brualla, Goldman, Nießner & Thies (2022) — persons/paper — flagged first, per the standing blind-spot: the real authors and real paper ("Neural RGB-D Surface Reconstruction," CVPR 2022) that the fabricated six-name Zhu et al. citation impersonates — the ground truth both Ansari's TF-coding and this capture's Claim 2 are measured against. "Azinović" is a known entity in the vault (3 mentions) with no entity page yet.
- Walters & Wilder (2023) — paper/concept — per Naser's own related-work section, the founding empirical study of LLM citation fabrication (*Scientific Reports*, GPT-3.5/GPT-4, 636 references) that both Ansari's and Naser's 2026 audits build on without either citing the other — the older ancestor underlying this whole cluster.
- GPTZero — organization/tool — already flagged in the 2026-09-20 capture as not-yet-checked for an entity page; resurfaces here as Ansari's sole data source (its automated Hallucination Check tool flagged all 100 citations she analyzed) and is directly thanked in her acknowledgments.
- M.Z. Naser — person — already has an entity page ([[entity-mz-naser]]); update not made this session since no new biographical fact surfaced, only a new claim about his paper's content.
- Samar Ansari — person — already has an entity page ([[entity-samar-ansari]]).
- Contamination Inheritance — concept/term — Ansari's own coinage; this capture adds that she flags it in her own Future Work section as potentially under-counted within her Total Fabrication category, a detail not yet in the vault's existing Ansari claim-notes.
