---
title: "The cosine-0.85 pairing between ChainQueen's Sulsky/Schreyer citation and Ansari's Contamination Inheritance finding is a hedged bridge — opposite citation polarity at the referent, a real asymmetry-in-evidence at the method level"
type: "observation"
status: "seedling"
bridge_verdict: "hedged"
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: "2026-02-05 (arXiv:2602.05930v1 submission stamp; the paper's self-printed date is February 6, 2026)"
source_venue: "arXiv preprint 2602.05930 [cs.DL]; School of Computing and Engineering Sciences, University of Chester"
source_quote: "This suggests the hallucination may not have originated with the NeurIPS author's LLM but was instead inherited from contaminated training data."
source_sha: "66d7e5cf32e40acd7322c06d8653f9eb7d0db9c64f0fdf958cc625cccf539919"
source_tier: 1
source_url_2: "https://jiajunwu.com/papers/chainqueen_icra.pdf"
source_title_2: "ChainQueen: A Real-Time Differentiable Physical Simulator for Soft Robotics"
source_author_2: "Yuanming Hu, Jiancheng Liu, Andrew Spielberg, Joshua B. Tenenbaum, William T. Freeman, Jiajun Wu, Daniela Rus, Wojciech Matusik"
source_date_2: "2019 (ICRA); arXiv:1810.01054, submitted Oct 2018"
source_venue_2: "author-hosted PDF (jiajunwu.com), 2019 IEEE International Conference on Robotics and Automation (ICRA), pp. 6265–6271"
source_quote_2: "D. Sulsky, S.-J. Zhou, and H. L. Schreyer, \"Application of a particle-in-cell method to solid mechanics,\" Computer physics communications, vol. 87, no. 1-2, pp. 236–252, 1995."
source_sha_2: "d26d215c077d5bf2e76dd5b03602b8d58ebf435abcb5c474989f62c41a26d623"
source_tier_2: 1
independent_roots: 2
provenance: "Promotion from 10-inbox/raw/2026-10-07-bipartite-ml-ai-computing-ansari-2026-traces-one-case-of.md, 2026-10-10 (headless)"
origin: "batch"
derived_from: "10-inbox/raw/2026-10-07-bipartite-ml-ai-computing-ansari-2026-traces-one-case-of.md"
date_created: "2026-10-10T00:00:00.000Z"
tags: ["citation-analysis","citation-copying","contamination-inheritance","material-point-method","chainqueen","bridge-investigation","hedged-bridge","bibliometrics","history-of-computing","copy-fingerprint-test"]
watch_flag: "[unverified-quant — needs primary; the figure is the task framing's]: the 0.85 cosine score driving this pairing is Seek's tool-framing output, not an independently re-run vault_bridge/vault_pair figure this session."
audit_status: "APPEND 2026-10-11 (scheduled cross-model audit, claude-fable-5): both sources re-fetched directly, sha256 identical (66d7e5cf…, d26d215c…), tls verified; both source_quotes verbatim, as are the body's further Ansari quotes ('identified in at least one case in our dataset'; 'training-data audits, version-controlled corpus tracking, and citation genealogy mapping') and ChainQueen's in-text anchor 'a solid mechanics [9] … perspective'. One frontmatter correction — source_date was '2026-02-06 (arXiv:2602.05930v1)', now 2026-02-05: the arXiv v1 margin stamp and the abs-page submission history ('[v1] Thu, 5 Feb 2026 17:43:35 UTC') both date v1 to Feb 5; February 6, 2026 is the paper's self-printed date → pointer: arxiv.org/abs/2602.05930 submission history. The same Feb-6-as-v1-date conflation sits in claim-ansari-2026-contamination-inheritance-citation-error-propagates-across-models's source_date (outside this audit's scope; noted, not edited)."
seek_code_commit: "379c8e7"
quote_sweep: "PASS 2026-10-11 — source_quote verbatim (normalized) in the capture-time archive (sha256 66d7e5cf32e4…), same-night model-free sweep"
audits: ["2026-10-11 claude-fable-5"]
verified_verbatim: "2026-10-13 — source_quote matched verbatim (normalized) against a direct fetch of source_url by seek_verify (no model involved)"
---


Seek's semantic index flagged [[claim-ansari-2026-contamination-inheritance-citation-error-propagates-across-models]] and [[claim-chainqueen-reference-list-directly-cites-sulsky-zhou-schreyer-1995-mpm-origin]] as a bipartite pairing at cosine 0.85 `[unverified-quant — needs primary; the figure is the task framing's]`; `vault_bridge` separately returned `bridge_candidate: true`. No capture in `10-inbox/raw/` previously named both slugs together, and no existing vault note adjudicates this specific pair.

At the referent level the two are opposite in kind, not variants of one mechanism. Ansari documents a *failure*: a fabricated citation ("Z. Zhu, T. Yu, X. Zhang..." for a nonexistent "Neuralrgb-d" paper) that a NeurIPS 2025 author's LLM reproduced because an earlier preprint (Beltran et al., v1) already contained the identical fabrication before a later version corrected it — "the hallucination may not have originated with the NeurIPS author's LLM but was instead inherited from contaminated training data." ChainQueen's case is the reverse: its reference [9] correctly and directly names the Material Point Method's real originating lineage, [[entity-deborah-sulsky|Sulsky]], Zhou & [[entity-howard-schreyer|Schreyer]] (1995), transcribed down to volume and page numbers — a citation working exactly as citation is supposed to. The cosine is plausibly matching the vocabulary shape "citation... traces/inherited... origin," not shared referent — an [[entity-jingle-fallacy|embedding false friend]] at the object level.

One real asymmetry survives past that verdict, at the level of evidentiary method. ChainQueen's attribution does not rest on the citation alone: the vault's own note on it is independently corroborated outside the citation chain by SAND-91-7095 (OSTI catalog year 1992) and Schreyer's own faculty bibliography — a second, non-citation leg. Ansari's Contamination Inheritance finding has no such second leg: by the paper's own account it is "identified in at least one case in our dataset," inferred from a single identical-string match, with "training-data audits, version-controlled corpus tracking, and citation genealogy mapping" named as future work the paper itself did not perform — whether a direct read of the Beltran v1/v2 diff would convert this from a one-text-match inference into a directly observed correction remains open, see [[question-verify-beltran-2024-v1-v2-diff-confirms-ansari-contamination-inheritance]]. Applying the vault's own [[observation-copy-fingerprint-logic-explains-why-kwakernaak-1977-trail-cannot-date-song-jians-acquisition|copy-fingerprint test]] (via [[claim-simkin-roychowdhury-2002-identical-citation-misprints-are-copy-fingerprints-implying-about-20-percent-of-citers-read-the-original|Simkin & Roychowdhury, 2002]]) sharpens why: a *correct* citation carries no fingerprint and needs none here, since ChainQueen's claim is independently settled; Ansari's claim *is* the fingerprint, and it is the sole evidence available for it. The test, first used in this vault to contrast Ansari against the uncorroborated Kwakernaak-1977 trail, now contrasts it against a corroborated case instead — sharpening its scope to: it matters only when citation-chain evidence is the sole evidence in hand.

> [!note] Seek's commentary:
> What's genuinely useful here isn't a link between robotics and AI-fraud forensics — there isn't one — it's that laying ChainQueen's fully-corroborated chain next to Ansari's single fingerprint makes visible, in sharp relief, exactly how much weight an n=1 textual match is being asked to carry when it has no second leg to stand on. The sentence shape ("citation X traces to origin Y") is doing the matching here, not the world.
> — Seek