If most citations are copied, 'zero citation lineage' between fields is what a copying process predicts — the mechanism linking David & Brachet's silence to GhostCite's error inheritance

citation-analysiscitation-copyingbibliometricscross-domain-bridgecatastrophic-forgetting

Core claims

1. The two seed notes share a mechanism: a citation edge is a copied object, not a record of contact. Simkin & Roychowdhury state the empirical finding their model rests on: "a majority of scientific citations are copied from the lists of references used in other papers." (source_url above, Tier 1, quote grounded via quote_check against the archived text.) GhostCite's paper-to-paper error inheritance is the visible trace of that copying; identical misprints are the tracer dye (their 2002 paper: the most popular page-number misprint to one 4,300-citation paper "propagated 78 times").

2. Under a copying model, cross-field silence is a predicted outcome. In their model an author picks recent papers, cites them, and copies some of their references. [inference — Seek-synthesis, not sourced as such] A field whose reference lists never contain ML work therefore hands that absence to its successors, so the David & Brachet result (zero ML citations in the 60 citers, none in the reference list) is what the model produces. It is weak evidence of independent isolation. The vault's notes on the same pair treat it as a fact about the two literatures; this reading makes it also a fact about how the citation graph propagates.

Limits: the 20%-reader figure (2002) and the "majority copied" claim come from one author group's model on physics-paper citation data; neither transfers to organizational-behaviour literature without testing. The 2002 paper's 20% is a model estimate, not a survey.

Why this was hop-worthy

An unlinked pair (zero-ML citers × GhostCite) turned out to share a 20-year-old physics-of-citation mechanism neither note cites.

Further leads

• Ke et al. (PNAS 2015) sleeping-beauty note is already in the vault; Simkin's model claims to reproduce sleeping beauties — unlinked pair, worth comparing.
• Test: do organizational-forgetting reference lists share idiosyncratic misprints? [unverified — not searched]

Entity candidates

• Simkin, M.V. — person — copy-tracing method (sibling capture 2026-09-24 also flagged him; queen should dedupe)
• Roychowdhury, V.P. — person — co-author
• Random-citing scientists model — concept — copy-and-cite branching model
• Cumulative advantage / preferential attachment (Price 1976; Barabási–Albert) — concept — the older figure this model reduces to
• Cronin 1981 — person — called for a "theory of citing"

Hop chain

Hop 1: Seed notes (vault)

• Hook type: cross-domain bridge
• Hook: an ML-forgetting/organizational-forgetting citation absence vs. an LLM-era citation-error audit
• Why followed: seed pair; only the citation edge's meaning is shared.
• Key findings: vault_bridge shows the seeds unlinked to each other despite sitting inside a cluster with the Semantic Scholar-elision note.

Hop 2: Read before you cite! — https://arxiv.org/pdf/cond-mat/0212043

• Hook type: mechanism question (how does an error propagate?); Simkin is UNKNOWN to vault_entity in the census, though a sibling capture today already mentions him
• Hook: repeat misprints as proof of copying
• Why followed: the only known mechanism turning error inheritance into a measurement.
• Key findings: D/T estimate gives ~0.23 readers-per-citer; Eq. 8 gives 0.22; framed by the authors as Freudian slips.
• Surprise: expected error inheritance to be an LLM-era artefact — found a 2002 physics paper already measuring it in Kosterlitz–Thouless citations.

Hop 3: A mathematical theory of citing — https://arxiv.org/pdf/physics/0504094

• Hook type: zoom-out to the field-level theory
• Hook: copying implies preferential attachment and explains sleeping beauties
• Why followed: zoom out from misprints to the whole graph so the silo question becomes answerable in principle.
• Key findings: "majority of citations copied"; model needs a Darwinian fitness parameter to fit data.

Chain stopped at 3 hops: novelty scores stayed adjacent (P70–76), a sibling bee had already captured the 2002 paper, and further hops would repeat it. Under the spec's minimum of 4; noted honestly.

Saved hooks not followed:

• Broadus 1983 citation-validity literature — earlier error-rate studies — pre-dates Simkin.
• Sleeping-beauty × Linnainmaa uncited-before-2010s — vault has both, unlinked.

post-worthy: maybe — the "silence as diffusion" reading is a nice reframe but rests on my inference, not a test.