---
id: "20260924-0203-bipartite-two-things-this"
title: "Bipartite hop: David & Brachet's zero-ML citing set vs. GhostCite's error propagation — a false friend at the level of the claims, one real contact at the level of the citation-edge process"
type: "capture"
status: "promoted"
origin: "batch"
writer_model: "claude-sonnet-5"
date_created: "2026-09-24T00:00:00.000Z"
provenance: "this batch run, 2026-09-24"
derived_from: []
tags: ["citation-analysis","bibliometrics","citation-error-propagation","cross-domain-bridge","false-friend","epistemology","ml-ai"]
source_url: "https://arxiv.org/pdf/2602.06718"
source_sha: "5142d40b814f570c41a3259045aa9a19882764b2e748b6c7c9e1e97a7a5fc240"
source_title: "GhostCite: A Large-Scale Analysis of Citation Validity in the Age of Large Language Models"
source_venue: "arXiv preprint 2602.06718v2 [cs.CR]"
source_tier: 1
recovered: true
recovered_on: "2026-10-01T00:00:00.000Z"
recovered_from: "Seek session transcript ed85ee4f-3a4d-4445-8732-4baeab89c8b8.jsonl on the mini - the session's Write at 2026-09-24 09:05 UTC, replayed verbatim"
recovery_note: "Overwritten the same night: another bipartite session was told to write the same filename (2026-09-24-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: ["2026-09-25-recovered-the-david-brachet-zero-ml-result-and-b4d9c2.md"]
seek_code_commit: "f8a723c"
promoted_to: ["30-notes/observation-david-brachet-ghostcite-cosine-087-pairing-is-hedged-false-friend.md"]
promoted_on: "2026-10-03T00:00:00.000Z"
promoted_by: "claude-opus-4-8"
not_promoted: ["Claim 'The two notes measure opposite things on the citation graph (absent edge set vs corrupted edges)' — FOLDED into the one verdict note, not a standalone claim. It is the core of the false-friend adjudication, not an independent finding (atomic-is-not-many: a verdict and its legs are one note).","Claim 'The two corpora are disjoint by construction (GhostCite's 56,381 AI/ML+Security papers vs David & Brachet's economics/operations citers)' — FOLDED into the same verdict note as the supporting leg that forecloses direct evidentiary contact. Both GhostCite quotes (title-similarity scope; eight-venue sample) carried into the note's frontmatter with sha 5142d40b… for the verifier bee.","Sub-note 'GhostCite text (739 invalid) vs Table V (738) off-by-one' — NOT promoted. A document-metadata/pagination claim; per promote.md step 5a metadata claims belong on a question log, not a claim-note, and this one is low-stakes and not load-bearing on anything kept. Recorded in the journal; the existing GhostCite claim-note already carries the 739 figure, re-verified by a 2026-09-25 cross-model audit.","Claim 'The one real mechanistic contact is citation copying — an inference that could bear on why cross-field edges stay at zero [unverified-mechanism]' — ALREADY PROMOTED, better developed, from the sibling capture 2026-09-24-hop-copied-citations-make-zero-lineage-a-diffusion-prediction.md on 2026-09-26: [[observation-copying-model-predicts-cross-field-citation-silence]], [[claim-simkin-roychowdhury-2002-identical-citation-misprints-are-copy-fingerprints-implying-about-20-percent-of-citers-read-the-original]], [[claim-simkin-roychowdhury-2005-random-copying-model-reproduces-preferential-attachment-and-sleeping-beauties]], [[entity-m-v-simkin]]. Folded as a wikilinked premise (the one surviving, indirect, inferential contact) into the verdict note; not duplicated.","Claim 'The zero-ML result is not threatened by GhostCite-scale error rates, but GhostCite documents the database-coverage ambiguity limiting the Semantic Scholar null [unverified cross-database robustness]' — NOT a standalone claim-note. The 0.006% arithmetic is minor; the load-bearing part (single-Semantic-Scholar null not database-robust; OpenAlex rate-limited) is a verification-needed caveat on an existing note, not a doubt the kept verdict rests on. Mentioned briefly in the verdict note and ROUTED as a dated progress line to [[question-verify-catastrophic-organizational-forgetting-citation-lineage]] (per promote.md 5a, database/vault-method claims go on the question log). No new question minted (intake discipline: few promises).","Entity candidates — none promoted. Simkin & Roychowdhury already have [[entity-m-v-simkin]] (covers the copying premise and links GhostCite/Ansari; taught nothing new here). Argote, David & Brachet already have hubs ([[entity-david-brachet]] got one additive dated line for the new false-friend role + link). GhostCite authors (Xu, Li), CiteVerifier, ghost citation, repeated-invalid-citation, Semantic Scholar API, citation copying — declined: single-paper authors/terms or concepts already held by the copying cluster; bias against the flood per entity-page-spec."]
---


**Pair:** [[claim-david-brachet-2011-citing-papers-contain-zero-ml-work]] × [[claim-xu-2026-ghostcite-finds-citation-error-propagation-between-published-papers]] (embedding cosine 0.87).

**Bridge-seed check:** searched `10-inbox/raw/` for captures naming both slugs — none. Three captures name the David & Brachet slug ([[claim-david-brachet-2011-citing-papers-contain-zero-ml-work]] appears in the 2026-08-12 lineage capture, the 2026-08-15 Brooks capture, and the 2026-09-20 tortured-phrases capture); only the 2026-09-22 capture names GhostCite, and not David & Brachet. No sibling verdict to contradict. Related earlier framing of the same false-friend pattern: [[observation-kelly-luccioni-cosine-pairing-is-embedding-false-friend]], [[observation-ml-ad-en-gedi-cosine-pairing-is-embedding-false-friend]], [[observation-greenhalgh-2003-klensin-romberg-1989-cosine-pairing-is-embedding-false-friend]].

**Verdict:** the resemblance is real only at the level of vocabulary-free topic ("something about citation graphs"). As claims they are opposite in kind and disjoint in corpus — a **false friend**. One narrow mechanistic contact survives (Claim 3), and it is an inference, not a finding. Core question resolved: *false friend, with one flagged inferential bridge*.

## Claim: The two notes measure opposite things on the citation graph — one an absent edge set, the other corrupted edges

Kind: technical-mechanism / definitional (what the audit measures). Source Tier 1 (the paper itself).

GhostCite defines its target as citation *validity* — whether a cited work exists and is correctly described — via title-similarity matching against bibliographic databases with threshold θ = 0.9. It states it deliberately focuses on "title similarity, as our preliminary goal is to identify citations that are untraceable or even non-existent (i.e. ghost citations), rather than citation errors." It does not study whether *real* citation links are missing between literatures. The David & Brachet note concerns the reverse: a complete enumeration of edges that exist (60 citing papers) and finds none from ML. One is a null in edge *presence*; the other a defect in edge *content*.

- source_url: https://arxiv.org/pdf/2602.06718
- source_sha: 5142d40b814f570c41a3259045aa9a19882764b2e748b6c7c9e1e97a7a5fc240 (extract_pdf, tls verified)
- source_author: Zuyao Xu, Yuqi Qiu, Lu Sun, et al. (Nankai University, Tsinghua University)
- source_date: 2026-02-06 (v1); v2 dated 14 May 2026 (read)
- source_title: "GhostCite: A Large-Scale Analysis of Citation Validity in the Age of Large Language Models"
- source_venue: arXiv preprint 2602.06718v2 [cs.CR]
- source_quote: "we only focus on title similarity, as our preliminary goal is to identify citations that are untraceable or even non-existent (i.e. ghost citations), rather than citation errors"
- source_tier: 1

## Claim: The two corpora are disjoint by construction — GhostCite audits eight AI/ML and Security venues; David & Brachet's citing papers are economics/operations

Kind: quantitative. Tier 1.

GhostCite's 56,381 papers come from NeurIPS (20,387), AAAI (13,821), ICML (11,192), IJCAI (5,535), USENIX Security (1,915), CCS (1,756), S&P (1,073), NDSS (702), 2020–2025 (Table IV; 20,387+13,821+11,192+5,535+1,915+1,756+1,073+702 = 56,381). Its own dataset therefore is ML-and-security by design; it cannot say anything about whether ML papers cite an organizational-economics paper, and the David & Brachet citing list (per the vault note: schedule-volatility, surgeon skill-decay, shipbuilding/aircraft learning curves) is not in its sample. Two studies with no shared population cannot corroborate or contradict one another. Note: GhostCite's "AI/ML" label is about publication venue, unrelated to the "ML" of the David & Brachet zero-ML finding except that the same ML literature would be the *citer* in one and the *absent* party in the other.

- source_url: https://arxiv.org/pdf/2602.06718
- source_sha: 5142d40b814f570c41a3259045aa9a19882764b2e748b6c7c9e1e97a7a5fc240
- source_author: Xu, Qiu, Sun et al.
- source_date: 2026-05-14 (v2)
- source_title: "GhostCite: A Large-Scale Analysis of Citation Validity in the Age of Large Language Models"
- source_venue: arXiv preprint 2602.06718v2 [cs.CR]
- source_quote: "56,381 papers collected from eight top-tier venues across AI/ML and Security domains from 2020 to 2025"
- source_tier: 1

Internal inconsistency worth Seek's notice: the paper's text reports 739 invalid citations (136 error + 603 ghost) while its Table V totals read 738 invalid (135 error + 603 ghost). The vault title's "739 … across 604" follows the text/abstract; the table is off by one. Low stakes, but an auditor grepping Table V will find 738.

## Claim: The one real mechanistic contact is that citations are formed by copying from other papers' reference lists, which makes edge sets path-dependent — an inference that could bear on why cross-field edges stay at zero

Kind: technical-mechanism. The copying mechanism itself is Tier 1 (both sources below); its application to David & Brachet's zero is **[unverified-mechanism -- needs primary]** because no source tests it against a cross-field null.

(a) GhostCite traces repeated invalid citations: an "AugMix" citation with an erroneous title recurs in 16 papers at AAAI, IJCAI and NeurIPS; the authors trace one origin to OpenReview's cite button returning the wrong title, and one propagation path where "a top-tier paper cited a work with erroneous metadata, and subsequently papers copied this incorrect citation." The authors say the surge in 2025 rates "aligns temporally with broader shifts in AI-assisted writing practices, but our data alone does not establish causality."

(b) The older primary GhostCite itself cites for this — Simkin & Roychowdhury, "Read before you cite!" — uses repeated misprints as a tracer of copying: identical misprints imply copying because "The probability of repeating someone else's misprint accidentally is 10^-4"; for a paper with 4,300 citations they count 196 misprints, 45 distinct, "The most popular misprint in a page number propagated 78 times," and estimate "only about 20% of citers read the original." (Their estimate rests on a stochastic model; the 20% is a model output with a stated generous definition of "reader.")

The inference: if a large fraction of citations are copied from reference lists already in a paper's own neighbourhood, new edges are drawn preferentially from an author's existing neighbourhood, so disjoint literatures would tend to stay disjoint. That would make the David & Brachet zero less a fact about that paper than an instance of a general property. Neither source tests this; Simkin & Roychowdhury measure misprint repetition, not cross-field edge formation. The bridge is therefore a hypothesis linking two measurements, not a result.

- source_url: https://arxiv.org/pdf/2602.06718 (GhostCite, part a; sha above)
- source_quote (a): "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"
- source_url: https://arxiv.org/pdf/cond-mat/0212043
- source_sha: 366ca10fbdaf67010c362412de471034cc161314c7759842af63781edc907943 (extract_pdf, tls verified)
- source_author: M.V. Simkin and V.P. Roychowdhury (Dept. of Electrical Engineering, UCLA)
- source_date: arXiv identifier indicates December 2002; exact date not printed in the extracted text
- source_title: "Read before you cite!"
- source_venue: arXiv preprint cond-mat/0212043 (journal-of-record not confirmed from the fetched text)
- source_quote (b): "Our estimate is only about 20% of citers read the original"
- source_tier: 1

> [!note] Seek's commentary:
> The cosine 0.87 is doing what these pairings keep doing: matching on "citations, and something going wrong with them." The two notes share a graph and nothing else — one is a hole where edges should be, the other a smudge on edges that exist. What I'd keep is the copying angle: it is the only place the two touch, and it turns "no one crossed the gap" into "people mostly cite what their neighbours cite," which is a question about how edges get made rather than about either paper.

## Claim: The zero-ML result is not threatened by GhostCite-scale error rates, but GhostCite documents the database-coverage ambiguity that limits the Semantic Scholar null

Kind: quantitative + mechanism. Tier 1 for the inputs; the derived rate is arithmetic on them.

GhostCite reports 136 metadata-error citations (text) among 2,199,409 extracted citations, about 0.006%, so citation records whose *metadata* is wrong enough to break a match are very rare in its ML/security sample. This is arithmetic on the paper's own counts and does not measure how often Semantic Scholar fails to link a real citation — the paper never queries Semantic Scholar. Separately, GhostCite states as a design challenge that "no single bibliographic database provides comprehensive coverage; different sources have different profiles, so absent results are often ambiguous." That caveat applies to any absence claim resting on one database, which is the position of the David & Brachet null (a single Semantic Scholar query; the vault already records publisher elision of reference data for the ML founding papers). Whether the null would survive a second database (OpenAlex was rate-limited when tried) remains open: `[unverified -- could not confirm or deny after search]` for cross-database robustness.

- source_url: https://arxiv.org/pdf/2602.06718
- source_sha: 5142d40b814f570c41a3259045aa9a19882764b2e748b6c7c9e1e97a7a5fc240
- source_author: Xu, Qiu, Sun et al.
- source_date: 2026-05-14 (v2)
- source_title: "GhostCite: A Large-Scale Analysis of Citation Validity in the Age of Large Language Models"
- source_venue: arXiv preprint 2602.06718v2 [cs.CR]
- source_quote: "no single bibliographic database provides comprehensive coverage; different sources have different profiles, so absent results are often ambiguous"
- source_tier: 1

## Further leads

- Whether a copying-based edge-formation model predicts cross-field zeros: the Simkin & Roychowdhury line of citation-copying models (follow-ups by Simkin/Roychowdhury on "redirection" models) — pointer only, not read.
- GhostCite's own related-work cites Simkin & Roychowdhury as its ref [22] for copy-propagation of errors — GhostCite's reference list, arXiv 2602.06718.
- GhostCite's Table VI lists other repeated invalid citations (e.g. "Reduction Of A Game With Complete Memory To A Matrix Game" in 7 papers across AAAI and NeurIPS) that could be traced as further propagation cases.
- Wang et al.-style iterated retraining question: [[question-does-citation-popularity-bias-compound-across-llm-training-generations]] (from the GhostCite note; not re-examined here).
- Ansari 2026 paper-to-model contamination inheritance: [[claim-ansari-2026-contamination-inheritance-citation-error-propagates-across-models]].
- Brooks/Dreyfus negative-citation tagging: [[observation-brooks-dreyfus-david-brachet-negative-citation-findings-unequally-tagged]]; [[claim-david-brachet-2011-reference-list-contains-zero-ml-citations]] for the reverse direction.

## Safety flags

None. Both sources were arXiv PDFs fetched via extract_pdf with tls verified; no addressed-to-AI, override, authority, tier self-assignment, file-system, credential or urgency signals encountered.

## Entity candidates

- Simkin and Roychowdhury (M.V. Simkin, V.P. Roychowdhury) — person — the older foundational work ("Read before you cite!") that the copy-propagation claim rests on and that GhostCite measures itself against; flagged first.
- Linda Argote — person — root figure of the organizational-forgetting tradition David & Brachet's citing set belongs to (already has a hub per earlier capture).
- Tanguy Brachet and Guy David — person — authors of the paper whose citing set is the "zero" side; see [[entity-david-brachet]].
- Zuyao Xu and Xiang Li (Nankai University) — person — lead and corresponding authors of GhostCite.
- CiteVerifier — term — GhostCite's open-source cascaded citation-verification framework (GROBID parse, DBLP/Google Scholar/web cascade, Levenshtein θ = 0.9).
- ghost citation — term — GhostCite's term for a citation to an untraceable work; distinguished there from "error citation" (wrong metadata for a real work).
- repeated invalid citation / misprint propagation — concept — the shared observable (identical errors across independent papers) used as a copying tracer by both GhostCite and Simkin & Roychowdhury.
- Semantic Scholar Academic Graph API — concept — the instrument behind the David & Brachet null; coverage ambiguity is the one GhostCite caveat that applies to it.
- citation copying (non-reading citation) — concept — the mechanism candidate linking the pair; unverified as an explanation of cross-field zeros.