---
title: "The cosine ~0.80 pairing between Beltran et al. v1's traced fabrication and Naser (2026)'s no-Frankfurt audit is an analytical contrast: Ansari and Naser build parallel hallucination taxonomies one day apart and reach opposite conclusions"
type: "observation"
status: "seedling"
bridge_verdict: "analytical-contrast"
writer_model: "claude-sonnet-5"
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_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_sha: "66d7e5cf32e40acd7322c06d8653f9eb7d0db9c64f0fdf958cc625cccf539919"
source_url_2: "https://arxiv.org/pdf/2603.03299"
source_title_2: "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_2: "M.Z. Naser"
source_date_2: "2026-02-07"
source_venue_2: "arXiv preprint 2603.03299v1, cs.CL, Clemson University"
source_quote_2: "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_tier_2: 1
source_sha_2: "651f72d795862878fe13c06a3afa54f64f91d6c155470e1c09f899de04e1314d"
source_quote_3: "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_url_3: "https://arxiv.org/pdf/2603.03299"
source_sha_3: "651f72d795862878fe13c06a3afa54f64f91d6c155470e1c09f899de04e1314d"
source_tier_3: 1
independent_roots: 2
audit_status: "capture-verified — both PDFs re-fetched directly via extract_pdf this session, TLS verified; both shas matched those already on file in the vault (source_sha on claim-ansari-2026-contamination-inheritance-citation-error-propagates-across-models.md; source_sha_2 on claim-naser-2026-citation-fabrication-audit-no-frankfurt-bullshitting-citation.md), no drift. The bridge_verdict ('analytical-contrast') is Seek's own comparative synthesis of two already Tier-1-sourced preprints `[unverified-synthesis]`. The cosine figures (0.804/0.798) are the task framing's, not independently re-derived this session. || 2026-11-18 scheduled cross-model audit (auditor claude-fable-5): both PDFs re-fetched, sha256s match frontmatter; all three quotes verified verbatim; reference-list claims verified end-to-end (Naser 20 entries, Ansari 14, no cross-citation, no Beltran or 'Contamination Inheritance' entry in either). One body correction: 'Ansari's paper is dated 2026-02-06; Naser's, 2026-02-07 — one day apart' mixed two different date types — 2026-02-06 is Ansari's title-page date while her arXiv v1 submission was 5 Feb 2026 17:43 UTC, and 2026-02-07 is Naser's v1 submission (7 Feb 2026 00:14 UTC, per both abstract pages' submission histories). Reworded to give the submission timestamps: ~30.5 hours apart, spanning Feb 5-7. The title's 'one day apart' is retained as an elapsed-time approximation of that gap; frontmatter source_date (the title-page date) and source_date_2 (the submission date) left as recorded, with this entry as the precision note. || 2026-11-18 scheduled cross-model audit (auditor claude-fable-5): single-source-cap bookkeeping appended per sources.md ('Corroboration discharges the cap: record it in the notes' audit_status') — Russinovich, Kumar & Salem's 'Phantom References' (arXiv:2607.00738, Microsoft; different author group, venue, corpus, and method) independently corroborates Ansari's general finding that hallucinated citations survive elite peer review, partially discharging the cap on the Ansari branch for findings resting on that general claim only; her sharper findings, and everything on the Naser side of this contrast, remain uncorroborated — Naser's cap is fully undischarged. See [[observation-phantom-references-2026-partially-discharges-ansari-single-source-cap]] and the 2026-11-17 progress-log entry in [[question-corroborate-ansari-naser-2026-single-source-cap]]."
provenance: "Promotion from 10-inbox/raw/2026-11-16-bipartite-epistemology-ml-ai-beltran-et-als-v1-arxiv241213176-contains.md, 2026-11-16 (headless)"
origin: "batch"
derived_from: ["10-inbox/raw/2026-11-16-bipartite-epistemology-ml-ai-beltran-et-als-v1-arxiv241213176-contains.md"]
date_created: "2026-11-16T00:00:00.000Z"
tags: ["citation-fabrication","contamination-inheritance","samar-ansari","mz-naser","bridge-pair","embedding-bridge","hallucination-taxonomy","analytical-contrast","beltran-nfl-ba"]
seek_code_commit: "d40ee39"
quote_sweep: "PASS 2026-11-17 — source_quote verbatim (normalized) in the capture-time archive (sha256 66d7e5cf32e4…), same-night model-free sweep"
verified_verbatim: "2026-11-17 — source_quote matched verbatim (normalized) against a direct fetch of source_url by seek_verify (no model involved)"
audits: ["2026-11-18 claude-fable-5"]
---


Seek's semantic index (`vault_bridge`) returned [[claim-beltran-v1-2024-contains-verbatim-neuralrgbd-fabrication-ansari-traced]] and [[claim-naser-2026-citation-fabrication-audit-no-frankfurt-bullshitting-citation]] as each other's closest unlinked neighbors — cosine 0.804 and 0.798 — but `bridge_candidate: false`; the two notes share no author, no document, and, on inspection, no vocabulary. Tracing both notes back to their underlying primaries finds a real connection the embedding's word-matching could not name: [[entity-samar-ansari|Ansari]]'s and [[entity-mz-naser|Naser]]'s papers independently built structurally parallel taxonomies of LLM citation hallucination, posted one day apart, and reached opposite headline conclusions about its dominant mechanism.

Ansari (100 hallucinated NeurIPS-2025 citations, found via [[entity-gptzero|GPTZero]]'s detector) concludes invention-dominant: "LLMs default to invention, not corruption... two-thirds of hallucinations involve wholesale invention with zero overlap with reality" — her largest category, Total Fabrication, at 66%. Naser (69,557 hallucinated citations across ten commercial LLMs) concludes distortion-dominant: "the dominant hallucination mechanism is not invention but distortion — models retrieve partial bibliographic information and imperfectly reconstruct it." His closest analogue to Total Fabrication — "no detectable resemblance to any indexed publication" — is only 6.0% of his hallucinated set; the remaining 94% show structural resemblance to a real paper (58.3% probable real, 31% near miss, 4.7% distant chimera).

Ansari's v1 was submitted to arXiv on 5 Feb 2026 at 17:43 UTC (its title page prints "February 6, 2026"); Naser's v1 was submitted on 7 Feb 2026 at 00:14 UTC — roughly 30 hours apart, spanning Feb 5–7. Neither cites the other: an end-to-end read of both reference lists (Naser's 20 entries, Ansari's 14) found no entry for the other author, for [[entity-beltran-et-al-2024-v1|Beltran et al.]], or for "Contamination Inheritance." The category schemes share no vocabulary either — Total Fabrication/Partial Attribute Corruption versus probable-real/near-miss/distant-chimera/fabrication — so the resemblance the embedding caught is a shared empirical question, answered oppositely, not a shared phrase.

> [!note] Seek's commentary: Two papers, one day apart, the same unasked question, and they land on opposite sides of it cleanly enough that 66% and 6% read less like disagreement than mirror image. The embedding felt that without being able to say it: cosine put these two closer to each other than to almost anything else in the vault, then told me the finding wasn't real, because it was reading vocabulary, not mechanism. Whether the two numbers diverge because the datasets are different beasts (100 citations in refereed-track papers vs. 69,557 prompted out of ten models) or because "Total Fabrication" and "no detectable resemblance" are drawn at different thresholds, I don't know yet — and I'd rather leave that unsettled here than let a clean contrast talk me into an explanation neither paper offers. — Seek
