Bipartite hop: Kwakernaak-1977 citation trail × Ansari-2026 Contamination Inheritance — a real mechanism (misprints as copy-tracers, Simkin & Roychowdhury 2002) that explains why the two cases are NOT the same kind of evidence
Pair: observation-kwakernaak-1977-citation-trail-sharpens-not-resolves-1975-dating-dispute × claim-ansari-2026-contamination-inheritance-citation-error-propagates-across-models (cosine 0.84).
Bridge-seed check: no capture in 10-inbox/raw/ names both slugs (checked by grep for the Ansari slug across the captures naming the Kwakernaak slug: zero overlap). The Ansari slug appears in five other captures (2026-09-13/14/20/22), none of which touch the Kwakernaak observation. Filename collision: the exact filename requested for this run (2026-09-24-bipartite-two-things-this-vault-knows-in-different.md) was already occupied by a sibling capture (id 20260924-0203, a different pair: David & Brachet × GhostCite). It was left untouched and this capture is written under a suffixed name.
Verdict: not a false friend of the vocabulary-free kind, and not an identity either. The shared mechanism is citation copying with identity-of-string as the only route-evidence — the same inference logic that lets Ansari call one error "inherited." Applying that logic to the Kwakernaak case shows why that trail cannot resolve the dating dispute: a correct citation carries no copy-fingerprint. The resemblance is a real mechanism, but it explains a difference between the two notes rather than a link that strengthens either.
Claim: Repeated identical citation misprints are the established evidence that citers copy references without reading the original
Kind: technical-mechanism (how copy-inference works) — Tier 1 required; source is the authors' own preprint (Tier 1).
Simkin & Roychowdhury argue that identical misprints cannot arise by coincidence and therefore mark copying from an earlier reference list. Their reasoning: for a 4-digit page number there are 10^4 possible misprints, so "the probability of repeating someone else's misprint accidentally is 10^−4," and "one concludes that repeat misprints are due to copying some one else's reference, without reading the paper in question."
- source_url: https://arxiv.org/pdf/cond-mat/0212043
- source_sha: 366ca10fbdaf67010c362412de471034cc161314c7759842af63781edc907943
- source_author: M.V. Simkin, V.P. Roychowdhury
- source_date: 2002-12
- source_title: "Read before you cite!"
- source_venue: arXiv preprint cond-mat/0212043
- source_quote: "There should be almost no repeat misprints by coincidence. One concludes that repeat misprints are due to copying some one else's reference, without reading the paper in question."
- source_tier: 1
Their data point (single "guinea pig" paper, Kosterlitz–Thouless 1973, 4300 citations): 196 misprints, 45 distinct, the most popular misprint "propagated 78 times"; estimated readers-to-citers ratio R ≈ 0.22–0.23 (Eq. 1 and 8), R ≈ 0.17 in the fuller model (Eq. 10); abstract: "Our estimate is only about 20% of citers read the original." Numbers are from the authors' own paper (Tier 1). They state that the distributions for "a dozen of other studied papers look very similar" (their footnote 10) — only one paper's data are shown in the preprint. The model is the authors' own, and this is an unrefereed preprint text as read; the "20%" is a model-dependent estimate, not a measurement.
Claim: Ansari's Contamination Inheritance case is inferred from the same fingerprint logic — an identical erroneous citation in an ancestor text — and Ansari flags it as n=1 with unknown prevalence
Kind: technical-mechanism / historical (what Ansari inferred). Tier 1 (the paper itself); the vault note claim-ansari-2026-contamination-inheritance-citation-error-propagates-across-models already records the quote; this record only adds the paper's own framing of what is and isn't established, read directly.
- source_url: https://arxiv.org/pdf/2602.05930
- source_sha: 66d7e5cf32e40acd7322c06d8653f9eb7d0db9c64f0fdf958cc625cccf539919
- source_author: Samar Ansari
- source_date: 2026-02-06
- source_title: "Compound Deception in Elite Peer Review: A Failure Mode Taxonomy of 100 Fabricated Citations at NeurIPS 2025"
- source_venue: arXiv preprint 2602.05930 [cs.DL], v1
- source_quote (§4.8.1): "Contamination Inheritance was identified in at least one case in our dataset: the Z. Zhu et al. citation discussed in Section 3.2, which traced back to an earlier arXiv preprint that was subsequently corrected."
- source_quote (same section): "Such investigations would reveal whether the contamination feedback loop, where AI-generated errors are learned by subsequent models, is already widespread or represents isolated cases."
- source_tier: 1
Two internal tensions worth carrying to promotion: (1) the same paper reports Total Fabrication as the primary mode for 66% of the 100 citations and says "LLMs default to invention, not corruption" (§4.3), so CI is a minority-or-unknown mechanism by the author's own accounting; (2) Ansari cannot observe the training data, so "inherited" is an inference from textual identity, not a demonstrated causal path — the paper itself lists "training-data audits" and "citation genealogy mapping" as the missing capabilities (§4.8.1). Whether Ansari's paper cites the misprint literature was not established in this read (unchecked).
Claim: The mechanism bridges the two notes only by contrast — the Kwakernaak trail lacks any copy-fingerprint, which is why it dates a printing but not an acquisition
Kind: analytic inference over the two existing vault notes and the two Tier 1 sources above; no new quantitative or mechanism claim beyond them. Marked [inference — Seek-synthesis, not sourced as such].
In the copy-inference method, evidence about the route of a citation comes from a shared error: identical misprints (Simkin & Roychowdhury) or an identical fabricated reference (Ansari). In the Song Jian case the evidence in hand (claim-greenhalgh-2003-article-documents-kwakernaak-1977-citation-in-song-1981, Greenhalgh 2003 p. 179, "Goldsmith et al. 1972; Kwakernaak 1977") is a correct citation. Nothing in a correct citation distinguishes "read at Twente in 1975 via Olsder," "copied from a Western reference list," or "read the original after 1975." So the trail bounds the latest date (in print by 1981) and cannot discriminate the channel — consistent with the existing observation's verdict that the citation "adds a third strand rather than resolving" the Song dating dispute, but here with a general reason rather than a case-specific one. The two notes are therefore the same epistemic problem (inferring provenance from citation edges) at opposite evidential strengths: Ansari has a fingerprint and n=1; the Kwakernaak trail has no fingerprint at all.
Limits: (a) This does not show Song's group copied the Kwakernaak citation — the vault notes and sources read here neither say nor exclude it; (b) an unusual feature (e.g. a shared idiosyncratic citation form in Kwakernaak-citing Chinese texts) would act as a tracer, but none has been examined ([unverified — not searched]); (c) the analogy is between a human bibliographic practice and an LLM training pipeline, and Simkin & Roychowdhury's 20% figure must not be transferred to either.
Central question status
Resolved as a partial, mechanism-level bridge by contrast (copy-inference via shared error). No claim that the Song Jian dispute has any causal or historical connection to LLM citation errors. The direct question — whether Song's Kwakernaak citation was copied or read — is [unverified — could not confirm or deny after search] (not searched beyond the vault's existing Greenhalgh note).
Further leads
- Simkin & Roychowdhury, follow-up "A mathematical theory of citing" (arXiv physics/0504094) and "Stochastic modeling of citation slips," Scientometrics (Springer link surfaced in search) — the "70–90% of citations copied" figure appeared only in a search summary; [unverified-quant — needs primary].
- The published venue of "Read before you cite!" (search summary says Complex Systems, 2003) — not fetched; confirm before citing a journal.
- Broadus (1983) on bibliographic-citation validity, and Simkin & Roychowdhury's own references — earlier error-rate literature; not searched.
- Ansari's Section 3.2 Zhu et al./Beltran et al. v1 (arXiv:2412.13176) — check the v1/v2 diff directly to confirm the ancestor really carries the identical string.
- claim-ml-and-ad-communities-mutually-unaware — the vault's existing note on disjoint literatures; relevant if citation-genealogy work in bibliometrics and in LLM-audit work turns out to be similarly unaware of each other.
- Sibling capture for a different pair with a partly overlapping theme:
2026-09-24-bipartite-two-things-this-vault-knows-in-different.md(David & Brachet × GhostCite).
Entity candidates
- Simkin, M.V. — person — first author of the repeat-misprint method; the foundational figure this capture measures Ansari's mechanism against
- Roychowdhury, V.P. — person — co-author, UCLA EE
- Kosterlitz–Thouless 1973 paper — term — the "guinea pig" paper with 4300 citations and 196 misprints
- Repeat misprint / copy-fingerprint inference — concept — using improbable identical errors to show citation copying
- Contamination Inheritance — concept — Ansari's named failure mode (already in vault claim)
- Citation genealogy mapping — concept — Ansari's proposed future-work method; formal cousin of misprint tracing
- Principle of Least Effort (Zipf) — concept — Simkin & Roychowdhury's explanation for low reader fraction
- Samar Ansari — person — author of the Ansari-2026 paper
- Susan Greenhalgh — person — source of the Kwakernaak-1977 citation evidence
- Huibert Kwakernaak — person — the cited Twente author
- Song Jian — person — the citing party in the dating dispute
- Geert-Jan Olsder — person — the 1975 account under dispute
- Beltran et al. (arXiv:2412.13176) — term — the contaminated ancestor preprint in Ansari's case