The Ansari Contamination-Inheritance × Chen, Liang & Wang (Naser-citing) cosine-0.888 pairing is an embedding false friend — fabricated vs. correct citation, no shared reference, chronology forecloses mutual citation
Seek's semantic index paired claim-ansari-2026-contamination-inheritance-citation-error-propagates-across-models against claim-chen-liang-wang-2026-cites-naser-as-bare-reference-not-empirical-test as a strong outlier — cosine 0.888 by direct re-measurement, the 4th rank in Ansari's own neighbour list and the 99.9th percentile of her cosine distribution against the vault's other 3,631 compared notes. The pair is unlinked in the graph, and no prior capture had investigated this specific bridge.
At the referent level the two claims describe opposite kinds of citation event. Ansari's Contamination Inheritance is a defective citation — a fabricated reference an LLM reproduced because it had already seen the same fabrication sitting, uncorrected, in an earlier contaminated preprint: "The language model likely encountered the erroneous citation in Beltran et al. (v1), learned it as a valid pattern, and reproduced it." Chen, Liang & Wang's citation of Naser is the reverse: a correct, intentional, undefective bibliographic reference in a refereed journal's related-work section, naming his real paper accurately by title and arXiv ID without testing his specific hallucination-rate finding — ordinary scholarly practice, not a failure mode.
A direct read of Chen, Liang & Wang's complete 39-entry reference list finds no entry for Ansari, for Beltran et al., or for the phrase "Contamination Inheritance" — only Naser, Xu et al.'s GhostCite, and Walters & Wilder (2023), the other members of the citation-fabrication-audit cluster Naser's branch descends from. Chronology forecloses the gap closing after the fact in the ordinary way, too: Ansari's single-version preprint (v1, "arXiv:2602.05930v1 [cs.DL] 5 Feb 2026") was frozen before Chen, Liang & Wang's paper existed ("Received: 02 April 2026... Published: 22 July 2026"), so it could not have cited a paper not yet written; Chen, Liang & Wang could in principle have cited Ansari's already-public preprint but did not.
This is the same shape the vault has now found for every other high-cosine pairing against Ansari's note: Ansari × Griewank resolved as a clean false friend on exactly this axis (fabricated vs. correct citation); Ansari × ChainQueen and Ansari × David & Brachet resolved as hedged bridges with a shared tool class this pairing does not share. The vocabulary the embedding keeps matching — "citation," "cites," "model," "finding/test" — recurs across this whole 2026 citation-integrity-audit cluster without tracking a shared mechanism, as David & Brachet × GhostCite already demonstrated for a different pair in the same cluster.
Source
“The language model likely encountered the erroneous citation in Beltran et al. (v1), learned it as a valid pattern, and reproduced it when generating references for computer vision topics.”
claude-sonnet-5 · Promotion from 10-inbox/raw/2026-11-19-bipartite-ml-ai-epistemology-ansari-2026-traces-one-case-of.md, 2026-11-19 (headless) · raw markdown