---
id: "20260925-0203-bipartite-two-things-this"
title: "The David & Brachet zero-ML result and GhostCite's error propagation share a copying mechanism only by inference; GhostCite's own evidence covers a different corpus and names a tool as a root cause"
type: "capture"
origin: "batch"
writer_model: "claude-sonnet-5"
date_created: "2026-09-25T00:00:00.000Z"
provenance: "batch run 2026-09-25, bipartite topic (epistemology × ml-ai); reads sibling capture 2026-09-24-hop-copied-citations-make-zero-lineage-a-diffusion-prediction.md"
derived_from: []
tags: ["citation-analysis","citation-copying","citation-error-propagation","false-friend","bridge-audit","cross-domain-bridge"]
source_url: "https://arxiv.org/pdf/2602.06718"
source_title: "GhostCite: A Large-Scale Analysis of Citation Validity in the Age of Large Language Models"
source_author: "Zuyao Xu, Yuqi Qiu, Lu Sun, Fasheng Miao, Fubin Wu, Xiang Li, et al. (Nankai University, Tsinghua University)"
source_date: "2026-05-14 (arXiv v2)"
source_venue: "arXiv preprint 2602.06718v2 [cs.CR]"
source_tier: 1
source_sha: "5142d40b814f570c41a3259045aa9a19882764b2e748b6c7c9e1e97a7a5fc240"
tls: "verified"
recovered: true
recovered_on: "2026-10-01T00:00:00.000Z"
recovered_from: "Seek session transcript b4d9c2f2-59b8-43f9-aba0-cc79977ea59d.jsonl on the mini - the session's Write at 2026-09-25 09:05 UTC, replayed verbatim"
recovery_note: "Overwritten the same night: another bipartite session was told to write the same filename (2026-09-25-bipartite-two-things-this-vault-knows-in-different.md). Collision fixed in seek-code c516ae8. Restored unpromoted at Cali's instruction, 2026-10-01. Before promoting, read the same-pair captures listed below: several sessions researched this pair, and only one per night survived."
same_pair_captures: ["2026-09-24-recovered-bipartite-hop-david-brachets-zero-ml-citing-ed85ee.md"]
seek_code_commit: "f8a723c"
status: "promoted"
promoted_to: ["30-notes/claim-ghostcite-augmix-propagation-root-cause-is-shared-tool-not-reader-copying.md"]
promoted_on: "2026-10-04T00:00:00.000Z"
promoted_by: "claude-sonnet-5"
not_promoted: ["Claim 'GhostCite's propagation finding rests on one traced path plus a repeat-count, and its authors hedge the inference' — FOLDED into the one promoted note rather than kept separate (atomic-is-not-many: a result and the caveat on how firmly it's known are one finding). Carried as the opening leg of [[claim-ghostcite-augmix-propagation-root-cause-is-shared-tool-not-reader-copying]].","Claim 'For the most-repeated error, GhostCite's own root-cause trace points to a tool, not to authors copying from each other's papers' — PROMOTED as [[claim-ghostcite-augmix-propagation-root-cause-is-shared-tool-not-reader-copying]], combined with the claim above into one note. Retrieve-before-write found no existing vault note made this specific common-cause-vs-chain distinction, though two notes (the false-friend verdict and the copying-model observation) use the same AugMix case as supporting evidence for the copying mechanism without noticing the distinction — the new note links to both and to [[entity-david-brachet]], which got one additive dated line.","Claim 'GhostCite's corpus is AI/ML and Security venues only, so it cannot test the cross-field claim' — NOT PROMOTED. This duplicates, almost sentence-for-sentence, two claims already promoted on 2026-10-03 from the same-pair sibling capture (2026-09-24-recovered-bipartite-hop-david-brachets-zero-ml-citing-ed85ee.md) into [[observation-david-brachet-ghostcite-cosine-087-pairing-is-hedged-false-friend]] (disjoint corpora; opposite kinds of measurement). Writing it again would be a duplicate, not a distinct angle.","Claim 'The copying explanation of cross-field silence is not discriminating, because independent topical selection predicts the same zero' — NOT PROMOTED. Explicitly marked in the capture as Seek-synthesis inference, not sourced as such, and the identical reasoning (including the 'cannot fingerprint an absence' framing) is already promoted and better developed at [[observation-copying-model-predicts-cross-field-citation-silence]], with the discriminating test already routed to [[question-do-organizational-forgetting-reference-lists-show-copy-fingerprints]]. A second note making the same point would duplicate an existing one rather than add to it.","Sub-note 'GhostCite's text (739 invalid) vs its own Table V (738) is off by one' — NOT PROMOTED. A document-internal-consistency/metadata claim (promote.md step 5a: metadata claims belong on a question log, not a claim-note), low-stakes, and not load-bearing on anything kept. The identical discrepancy was found and the identical decision made from the same-pair sibling capture on 2026-10-03; no new question minted, consistent with that precedent. [[claim-xu-2026-ghostcite-finds-citation-error-propagation-between-published-papers]] already carries the 739 figure, re-verified by a 2026-09-25 cross-model audit.","Entity candidates — none promoted to a new hub or stub. Simkin and Roychowdhury already have [[entity-m-v-simkin]] (no new fact to append). Cronin (1981) and Price de S. (1976) are cited in this capture only 'per the sibling capture' — secondhand, not independently verified here — declined on that basis. Zuyao Xu, CiteVerifier, Ghost citation, Repeated invalid citation, Epistemic pollution, OpenReview cite button, and Google Scholar 'Cite' button are single-paper terms or narrow tool-specifics already folded into the promoted claim-note's body rather than stood up as watching stubs; none is recurring enough yet to clear the bias-against-the-flood bar. McCloskey & Cohen already has [[entity-mccloskey-cohen]] with nothing new taught here. David & Brachet already has [[entity-david-brachet]] as a hub, updated additively this session with the new claim's one-line finding, per the 'already has a page' discipline."]
---


## Sibling check (bridge-seed discipline)

A sibling capture already investigated this exact pair: `10-inbox/raw/2026-09-24-hop-copied-citations-make-zero-lineage-a-diffusion-prediction.md`. It proposes that a copying process (Simkin & Roychowdhury) links the two notes, so that "zero citation lineage" between fields is what copying predicts, and rates the link "maybe — rests on my inference, not a test." A second sibling, `2026-09-22-has-anyone-run-the-wang-et-al-style.md`, is the origin capture of the GhostCite note and is not a bridge investigation. This capture does not re-derive Simkin's model. It re-reads GhostCite, which the sibling cited only through the vault note, and records where that read **partly disagrees** with the sibling's verdict.

**Verdict of this capture:** a real but thin bridge, not a false friend. The two notes share the object "citation edge" and both are consistent with citation copying. GhostCite's own text does not support extending the mechanism to cross-field silence, and the extension makes no prediction that independent topical selection does not also make. Central question (does copying causally link the two findings?): **[unverified -- could not confirm or deny after search]**.

## Claim: GhostCite's "propagation" finding rests on one traced path plus a repeat-count, and its authors hedge the inference

*Kind: technical-mechanism + quantitative. Source is a Tier 1 primary (the paper itself), so it clears the floor.*

GhostCite reports that the erroneous citation of AugMix appears in 16 papers across AAAI, IJCAI and NeurIPS. The authors say they "traced one clear propagation path where a top-tier paper cited a work with erroneous metadata, and subsequently papers copied this incorrect citation." For the wider pattern they hedge: "These viral errors suggest that researchers may be copying citations from other papers that already contain errors." So copying is a stated hypothesis, supported by one traced chain and a repeat-count table (Table VI, top entry 16 papers, next entry 7).

- source_url: https://arxiv.org/pdf/2602.06718
- source_title: GhostCite: A Large-Scale Analysis of Citation Validity in the Age of Large Language Models
- source_author: Xu, Qiu, Sun et al.
- source_date: 2026-05-14 (v2)
- source_venue: arXiv preprint 2602.06718v2 [cs.CR]
- source_sha: 5142d40b814f570c41a3259045aa9a19882764b2e748b6c7c9e1e97a7a5fc240
- source_tier: 1
- source_quote: "We even traced one clear propagation path where a top-tier paper cited a work with erroneous metadata, and subsequently papers copied this incorrect citation, continuing to propagate the mistake across the literature."
- source_quote_2: "These viral errors suggest that researchers may be copying citations from other papers that already contain errors, compounding mistakes across the literature."

Relation to the existing note [[claim-xu-2026-ghostcite-finds-citation-error-propagation-between-published-papers]]: the note states propagation as an established finding. The paper states it as observed for one path and suggested for the rest.

## Claim: For the most-repeated error, GhostCite's own root-cause trace points to a tool, not to authors copying from each other's papers

*Kind: technical-mechanism. Tier 1 primary.*

For the 16-paper AugMix error, the authors write that "we found that OpenReview's cite button returns this incorrect title, which may be the source of this repeated error." This is a common upstream source, and it is a different mechanism from paper-to-paper copying: 16 authors can carry the same error without any of them reading another's reference list. The survey figures point the same way: 72.6% of respondents rely on Google Scholar's "Cite" button and 4.8% copy BibTeX entries from other papers [numbers from the same paper's Section VII; survey n=94, self-reported]. Repeat counts alone therefore cannot separate the two mechanisms. This weakens the use of GhostCite as a demonstration of the paper-to-paper diffusion that the sibling's bridge needs.

- source_url: https://arxiv.org/pdf/2602.06718
- source_title: GhostCite: A Large-Scale Analysis of Citation Validity in the Age of Large Language Models
- source_sha: 5142d40b814f570c41a3259045aa9a19882764b2e748b6c7c9e1e97a7a5fc240
- source_tier: 1
- source_quote: "we found that OpenReview's cite button returns this incorrect title, which may be the source of this repeated error."

> [!note] Seek's commentary:
> This is the disagreement with the sibling capture. The sibling reads GhostCite's inheritance as the visible trace of the copying process that also produces the silo. In GhostCite's own data the flagship repeat has a shared tool upstream. That is a diffusion story of a different shape, closer to a common cause than a chain.

## Claim: GhostCite's corpus is AI/ML and Security venues only, so it observes no organizational-behaviour literature and cannot test the cross-field claim

*Kind: quantitative/scope. Tier 1 primary.*

The audit covers "papers at AI/ML and Security venues (2020–2025)": NeurIPS, ICML, AAAI, IJCAI, USENIX Security, CCS, S&P, NDSS. The David & Brachet citing set (60 papers on organizational forgetting, skill decay and learning curves, per the vault note) has no overlap with that corpus by construction. GhostCite measures **invalid citations** (untraceable or mis-titled entries). David & Brachet's result is an **absence of edges** to ML work, and every edge in it is a valid one. The two datasets differ in unit (error vs. absence), field, and time window, so they can only be joined through a theory, not through data. GhostCite also warns against causal reading of its own trend: "our data alone does not establish causality."

- source_url: https://arxiv.org/pdf/2602.06718
- source_title: GhostCite: A Large-Scale Analysis of Citation Validity in the Age of Large Language Models
- source_sha: 5142d40b814f570c41a3259045aa9a19882764b2e748b6c7c9e1e97a7a5fc240
- source_tier: 1
- source_quote: "papers at AI/ML and Security venues (2020–2025)"
- source_quote_2: "our data alone does not establish causality"

Related vault notes: [[claim-david-brachet-2011-citing-papers-contain-zero-ml-work]], [[entity-david-brachet]], [[observation-brooks-dreyfus-david-brachet-negative-citation-findings-unequally-tagged]].

## Claim: The copying explanation of cross-field silence is not discriminating, because independent topical selection predicts the same zero

*Kind: inference (Seek-synthesis, not sourced as such; no factual figure is asserted).*

The sibling argues that if authors copy references from the papers they read, a field with no ML in its lists passes that absence on. The competing null is that org-forgetting authors cite org-forgetting work because it is on topic, and ML work is off topic for them. Both predict zero ML citers among the 60. The copying reading would only be distinguishable if it predicted something the null does not, for example shared idiosyncratic misprints, or ML papers that are topically relevant yet absent from lists that copy from each other. Neither has been tested [unverified -- not searched for, no data located]. A related caution: GhostCite itself cites Simkin & Roychowdhury for "a substantial fraction of citations are copied from secondary sources," but describes this as prior estimate, and the sibling notes that the model comes from one author group's work on physics citations.

> [!note] Seek's commentary:
> Framed this way, the two findings are consistent with each other, not evidence for each other. The bridge earns a link (both notes concern what a citation edge does and does not certify) but it does not earn the word "mechanism" until a copy-tracer is run on the organizational-forgetting cluster.

## Internal discrepancy noted in GhostCite (for the existing claim-note)

GhostCite's text reports "739 (29.2%) were confirmed as invalid (136 error citations and 603 ghost citations)," while its Table V totals read Error 135, Ghost 603, Invalid 738. The vault note's headline figure of 739 across 604 papers matches the abstract and the running text, but the paper's own table sums to 738. The discrepancy is one citation and is unexplained in the paper. Both figures are Tier 1 from the same document.

- source_url: https://arxiv.org/pdf/2602.06718
- source_sha: 5142d40b814f570c41a3259045aa9a19882764b2e748b6c7c9e1e97a7a5fc240
- source_quote: "739 (29.2%) were con-firmed as invalid (136 error citations and 603 ghost citations)" — text is hyphen-broken across a column line in the cached extraction ("con-" / "firmed"); grep for "29.2%" to locate.

## Further leads

- Simkin & Roychowdhury, "Read before you cite!" (arXiv cond-mat/0212043) and "A mathematical theory of citing" (arXiv physics/0504094): sibling capture holds the sha256 receipts; not re-read this session.
- GhostCite's Table VI includes classic ML titles among repeated invalid citations (e.g. "Keeping Neural Networks Simple By Minimising The Description Length Of The Weights", 3 papers) — a copy-tracer already sits in the data for anyone wanting an ML-side propagation case; source: same paper, Table VI.
- Ansari 2026 "Contamination Inheritance" (paper-to-model propagation), already in the vault as [[claim-ansari-2026-contamination-inheritance-citation-error-propagates-across-models]]-style note, per the GhostCite note; not re-checked here.
- Testing the discriminating prediction: do the org-forgetting citers' reference lists share idiosyncratic misprints? Semantic Scholar's citation list for David & Brachet is the entry point (source: existing vault note's API URL).
- Related vault observations on embedding false-friends for method comparison: [[observation-ach-abduction-lucky-widrow-cosine-pairing-is-embedding-false-friend]], [[observation-oaksford-chater-wason-resource-rational-pairing-is-confirmed-bridge-via-citation]] (a confirmed bridge via an actual citation, the contrast case).

## Entity candidates

- Simkin, M.V. — person — the copy-tracing/misprint method; GhostCite itself cites him; the foundational figure the mechanism claim rests on
- Roychowdhury, V.P. — person — co-author of the copying model
- Cronin, B. (1981) — person — earlier call for a theory of citing, per the sibling capture
- Price, D. de S. (1976) — person — cumulative-advantage model that copying reduces to, per the sibling capture
- Zuyao Xu — person — GhostCite first author
- CiteVerifier — term — GhostCite's open-source verification pipeline
- Ghost citation — term — GhostCite's definition: untraceable citation, distinct from a metadata-error citation
- Repeated invalid citation — concept — same wrong reference in multiple independent papers; the measurable trace of propagation
- OpenReview cite button — term — named by GhostCite as a probable common upstream source of the AugMix error
- Epistemic pollution — concept — GhostCite's phrase for non-existent studies cited to support claims
- Google Scholar "Cite" button — term — used by 72.6% of survey respondents for references; a shared upstream source
- McCloskey & Cohen (1989) — person — the ML side of the catastrophic-forgetting lineage whose absence from org-forgetting lists is the underlying seed question
- David & Brachet — person — the seed paper's authors, the older figure the zero-ML claim is measured against (vault entity exists: [[entity-david-brachet]])