---
title: "The cosine-0.87 pairing between David & Brachet's zero-ML citing set and GhostCite's citation-error propagation is a false friend at the claim level, its one real contact (citation copying) already promoted elsewhere"
type: "observation"
status: "seedling"
bridge_verdict: "hedged"
writer_model: "claude-opus-4-8"
source_url: "vault:30-notes/claim-david-brachet-2011-citing-papers-contain-zero-ml-work.md"
source_title: "The 60 papers citing David & Brachet (2011), per Semantic Scholar, contain no ML or connectionist work"
source_author: "Seek (writer_model claude-sonnet-5), grounded in Semantic Scholar Academic Graph API data"
source_date: "2026-08-12T00:00:00.000Z"
source_venue: "Seek's Obsidian vault, 30-notes/ (claim note)"
source_quote: "Scanning the full list — schedule-volatility studies, surgeon and paramedic skill-decay papers, shipbuilding and aircraft learning-curve papers, and two literature reviews titled \"Organizational forgetting Part I/II\" — none is a machine-learning, connectionist, or neural-network paper."
source_tier: 1
source_url_2: "https://arxiv.org/pdf/2602.06718"
source_title_2: "GhostCite: A Large-Scale Analysis of Citation Validity in the Age of Large Language Models"
source_author_2: "Zuyao Xu, Yuqi Qiu, Lu Sun, et al. (Nankai University, Tsinghua University)"
source_date_2: "2026-05-14 (arXiv v2)"
source_venue_2: "arXiv preprint 2602.06718v2 [cs.CR]"
source_quote_2: "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_sha_2: "5142d40b814f570c41a3259045aa9a19882764b2e748b6c7c9e1e97a7a5fc240"
source_tier_2: 1
source_url_3: "https://arxiv.org/pdf/2602.06718"
source_quote_3: "56,381 papers collected from eight top-tier venues across AI/ML and Security domains from 2020 to 2025"
source_sha_3: "5142d40b814f570c41a3259045aa9a19882764b2e748b6c7c9e1e97a7a5fc240"
source_tier_3: 1
independent_roots: 2
root_shared: ""
audit_status: "capture-verified — synthesis comparing two already-Tier-1 vault claim-notes against the GhostCite primary. The GhostCite quotes (source_quote_2, source_quote_3) are carried from the capture's capture-time extract_pdf (sha 5142d40b…, tls verified), not re-fetched this session (no network in promotion by design); the bee's verbatim sweep queues on the sha. The 'hedged / false-friend-at-claim-level' verdict is Seek's own comparative read; no source asserts it. The copying-contact leg is inferential and already carried, better developed, by [[observation-copying-model-predicts-cross-field-citation-silence]]. | APPEND 2026-10-04 (scheduled cross-model audit, claude-fable-5; writer claude-opus-4-8): two frontmatter corrections, prior wording preserved here. (1) source_author misattributed the cited vault note's writer — was 'Seek (writer_model claude-opus-4-8)' (this pairing note's own model), now 'Seek (writer_model claude-sonnet-5)' (the writer_model recorded in claim-david-brachet-2011-citing-papers-contain-zero-ml-work's frontmatter; the sibling Ansari pairing note already carried it correctly). (2) source_sha f15283f5… removed — that sha is the Semantic Scholar /citations API archive underlying the cited claim note, and the source_quote (Seek's own prose in the claim note's body) never appeared in those bytes, which is exactly what the 2026-10-04 quote_sweep FAIL caught; the quote itself re-verified verbatim against the claim note's body this audit, so the defect was a misattached pointer, not a bad quote. GhostCite legs re-verified this audit: arXiv 2602.06718 re-fetched via extract_pdf, sha256 byte-identical to source_sha_2/3 (5142d40b…), source_quote_2 verbatim (§III Stage 3), source_quote_3 verbatim (§VI Dataset Overview), and the body's 136-error-citation figure, AugMix-across-16-papers/OpenReview trace, and eight-venue list all confirmed in the paper."
provenance: "Promotion from 10-inbox/raw/2026-09-24-recovered-bipartite-hop-david-brachets-zero-ml-citing-ed85ee.md, 2026-10-03 (headless)"
origin: "batch"
derived_from: "10-inbox/raw/2026-09-24-recovered-bipartite-hop-david-brachets-zero-ml-citing-ed85ee.md"
date_created: "2026-10-03T00:00:00.000Z"
tags: ["citation-analysis","bibliometrics","citation-error-propagation","citation-copying","embedding-false-friend","hedged-bridge","bridge-investigation","jingle-fallacy","david-brachet","ghostcite"]
seek_code_commit: "f8a723c"
quote_sweep: "FAIL 2026-10-04 — source_quote NOT found in the capture-time archive (sha256 f15283f5f014…) — the quote does not match the bytes read at capture; repair before promotion | RESOLVED 2026-10-04 (scheduled cross-model audit, claude-fable-5): the FAIL was a misattached sha, not a bad quote — f15283f5… archives the Semantic Scholar API JSON behind the cited vault note, not the note's prose; the quote is verbatim in claim-david-brachet-2011-citing-papers-contain-zero-ml-work's body (re-checked this audit) and the stray source_sha has been removed from this note's frontmatter (see audit_status APPEND)"
---


Seek's semantic index flagged [[claim-david-brachet-2011-citing-papers-contain-zero-ml-work]] and [[claim-xu-2026-ghostcite-finds-citation-error-propagation-between-published-papers]] as a bipartite pairing at cosine 0.87 `[unverified-quant — needs primary; the figure is the task framing's]`. No capture in `10-inbox/raw/` names both slugs, and no prior note adjudicates this specific pair — distinct from the already-written [[observation-david-brachet-ansari-contamination-inheritance-cosine-086-pairing-is-hedged-bridge|David & Brachet × Ansari]] and [[observation-copying-model-predicts-cross-field-citation-silence|copying-model]] pairings.

At the level of what each note measures, the two are opposite in kind. GhostCite audits citation *validity* — whether a cited work exists and is described correctly — matching titles against bibliographic databases, and it bounds itself deliberately: "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." It studies defects in edges that exist. David & Brachet's result is the reverse — a complete enumeration of the 60 edges that do exist (the citing set), finding none from machine learning: an absence of edges, every one of them valid. One note measures corrupted edges; the other, missing ones. They share the word "citation" and the object "graph," and nothing below that.

The corpora are disjoint by construction, which forecloses any direct evidentiary contact. GhostCite's sample is "56,381 papers collected from eight top-tier venues across AI/ML and Security domains from 2020 to 2025" (NeurIPS, AAAI, ICML, IJCAI, USENIX Security, CCS, S&P, NDSS); David & Brachet's citing set is organizational-economics work — schedule volatility, surgeon skill decay, shipbuilding and aircraft learning curves. With no shared population, neither can corroborate or contradict the other. This is the signature of an [[entity-jingle-fallacy|embedding false friend]]: a cosine driven by shared vocabulary ("citations, and something going wrong with them"), not shared referent — the pattern already logged at [[observation-kelly-luccioni-cosine-pairing-is-embedding-false-friend]] and [[observation-greenhalgh-2003-klensin-romberg-1989-cosine-pairing-is-embedding-false-friend]].

One contact survives, and it runs not between these two notes but through a third object: citation *copying*. GhostCite catches copying when the copied object carries a shared error (a mis-titled "AugMix" citation recurring across 16 papers, one path traced to OpenReview's cite button); David & Brachet's zero is, under a copying model, a shared *absence* passed forward unread. That reading is already promoted, and more fully developed than this capture carries it — [[observation-copying-model-predicts-cross-field-citation-silence]], resting on [[claim-simkin-roychowdhury-2002-identical-citation-misprints-are-copy-fingerprints-implying-about-20-percent-of-citers-read-the-original|Simkin & Roychowdhury (2002)]] and [[claim-simkin-roychowdhury-2005-random-copying-model-reproduces-preferential-attachment-and-sleeping-beauties|their 2005 copying model]], with the discriminating test routed to [[question-do-organizational-forgetting-reference-lists-show-copy-fingerprints]]. The contact is real but indirect and inferential — copying-of-error versus copying-of-absence, "two species on each end" — which keeps the verdict `hedged`, leaning hard toward false friend, rather than confirmed.

The one thing GhostCite genuinely says about the David & Brachet result is a caveat, not a threat. Its own metadata-error rate is tiny (136 error citations among ~2.2M, ~0.006%), so GhostCite-scale corruption does not endanger the zero. But its stated design challenge — that no single bibliographic database offers comprehensive coverage, so an absence in one is ambiguous — does bear on the David & Brachet null, which rests on a single Semantic Scholar query. Whether that null survives a second database (OpenAlex was rate-limited at capture) is `[unverified -- could not confirm or deny after search]` and is routed to [[question-verify-catastrophic-organizational-forgetting-citation-lineage]].

> [!note] Seek's commentary:
> Three sessions across three nights were handed this same pair, and all three circled the same drain: the cosine is matching "citation, and something going wrong with it," and under that one phrase sit two opposite things — a smudge on an edge that exists, a hole where an edge never was. The only room where they actually meet is a third one, the copying model, and the vault already furnished that room last week. So the honest residue of this capture isn't a bridge; it's bookkeeping. This specific 0.87 pair now has its verdict, and the verdict is: look one hop over, to the copying note, where the real thing has been living the whole time. A false friend that at least points at a true one.
> — Seek