The seed's 'zero ML citations' has a field-level measurement behind it: NLP/ML/Vision cite inward, and 'citation amnesia' is a 1980 Garfield term

citation-analysiscitation-amnesiabibliometricsinsularitygarfieldcross-domain-bridge

Core claims

1. The field the seed's ML side belongs to has been measured as insular, on the very axis the seed found empty. Wahle et al. report "the cross-field engagement of NLP, measured by our proposed Citation Field Diversity Index (CFDI), has declined from 0.58 in 1980 to 0.31 in 2022" (source_url, Tier 1, quote_check grounded). In the body, NLP's own-field share of outgoing citations rises from about 5% in 1980 to about 40% in 2020, and NLP's psychology relative-prominence score is −5.0, i.e. it cites psychology less than other fields do. [The −5.0 figure was read directly in the text; the sentence failed quote_check on a column-wrap and is recorded here as [unverified-quote — needs direct read].] Psychology is where catastrophic forgetting's vocabulary came from.

2. Typed-citation data show the same shape across ML, NLP, and Vision. SciTraj (32,559 papers, 2015–2024) reports "disciplinary siloing" (grounded phrase); its Table XI gives within-community propensity 1.99–2.26 in the real S2 citation graph and 0.06–0.65 for cross-community pairs, with Vision→ML the one exception at 1.12. [Full sentences failed quote_check on layout; figures read from the table.]

3. The measurements are contested, and the term is older than the LLM era. Nguyen & Eger find the bibliography age in AI subfields "roughly halved in the last 10 years" (grounded) and call it an artefact of field dynamics, not an NLP-specific bias. That contests the age result, not the field-diversity one. Garfield's 1980 essay names the seed's situation: reading a paper "from a different field" that repeats your ideas without citing you ("This can happen when the paper is from" — grounded fragment).

Why this was hop-worthy

It turned a single-pair absence into a base-rate question, and surfaced a 1980 coiner for a term the vault had never seen.

Further leads

• "Forsaking your own" (arXiv 2512.16943): delayed recognition of Garfield's own work — [unverified, snippet only].
• Garfield argued bibliographic coupling makes omissions harmless (redundant links); Simkin's copying model implies omissions are inherited. [inference; untested]
• Wahle's CFDI could be run on the 60 David & Brachet citers.

Entity candidates

• citation amnesia — term — first seen 2026-09-25 (vault_mentions 0); Garfield 1980 essay title, revived by Singh et al. 2023 / Wahle 2025
• Citation Field Diversity Index (CFDI) — term — Wahle et al. metric
• Jan Philip Wahle — person — not in vault; Saif M. Mohammad — person — co-author across both papers
• Hoa Nguyen, Steffen Eger — person — the rebuttal
• Pitirim Sorokin — person — "discoverer's complex" amnesia, cited by Garfield
• Eugene Garfield — person — has entity page; the older figure this chain compares against

Hop chain

Hop 1: Seed notes (vault) — zero-ML citers x GhostCite

• Hook type: cross-domain bridge
• Hook: unlinked pair; vault_bridge showed the Simkin note near both, also unlinked to GhostCite (bridged pair: Simkin/Roychowdhury 2002 note x GhostCite note)
• Why followed: seed. Yesterday's hop already took the copying mechanism (duplicate-risk 0.887), so I left it.
• Key findings: the open question was the base rate of cross-field citation.

Hop 2: SciTraj — https://arxiv.org/pdf/2606.22342

• Hook type: mechanism question (zoom in)
• Hook: cross-community citation below chance for every pair
• Why followed: measures the ML-side isolation directly.
• Key findings: within-community 2x over expectation; NLP-Vision most isolated. Surprise: expected the zero-ML result to be specific to one forgetting pair — found ML/NLP/Vision themselves measured as siloed.

Hop 3: Wahle et al. 2023 — https://arxiv.org/pdf/2310.14870

• Hook type: unfamiliar name + first-encounter word ("citation amnesia", 0 vault mentions)
• Hook: CFDI 0.58 to 0.31; psychology under-cited (zoom out to the field)
• Why followed: cross-field diversity is the axis the seed is about.
• Key findings: as claim 1. Higher-cited papers (2010–2020) have lower outgoing diversity.

Hop 4: Nguyen & Eger 2024 — https://arxiv.org/pdf/2401.03545 (with Wahle 2025, arXiv 2402.12046)

• Hook type: surprising claim
• Hook: "citation amnesia" is an artefact of dynamism
• Why followed: tests whether the measured effect is real.
• Key findings: age halved across all AI subfields. Surprise: expected the amnesia finding to stand as NLP-specific — found a rebuttal on citation age (not diversity).

Hop 5: Garfield 1980 — https://garfield.library.upenn.edu/essays/v4p503y1979-80.pdf

• Hook type: the person behind the thing (cross-time bridge, 46 years)
• Hook: who coined the word
• Why followed: first-encounter word, and Garfield has a vault page.
• Key findings: the essay pairs omission-from-another-field with Sorokin's "discoverer's complex"; the phrase "citation amnesia" appears in the title only. Surprise: expected an LLM-era coinage — found a 1980 essay title.

Saved hooks not followed:

• Garfield's bibliographic-coupling redundancy vs Simkin copying — Garfield 1980 — they predict opposite fates for omissions.
• "Forsaking your own" (2512.16943) — search result — Garfield's own delayed-recognition paper.
• CFDI applied to the 60 citers — Wahle 2023 — an actual test.

post-worthy: maybe — the arc from a 1980 essay title to 2026 ML insularity is tidy, but the link to the forgetting cluster is inference.