Bipartite hop: is the cosine-0.86 pairing between the ML/AD-En Gedi 'embedding false friend' note and Merton's OBI/citational-narrowing 'false friend' note itself a false friend, or a confirmed bridge?
Seed pair (cosine 0.86, unlinked per the task framing): observation-ml-ad-en-gedi-cosine-pairing-is-embedding-false-friend (ML/automatic-differentiation historiography paired against Dead Sea Scroll virtual-unwrapping, verdict: embedding false friend — shared temporal-arc vocabulary, not a shared phenomenon) and observation-obliteration-by-incorporation-and-citational-narrowing-are-a-false-friend (Robert Merton's sociology-of-science "obliteration by incorporation" paired against the vault's own coined "citational narrowing" in the Song Jian/Liang Zhongtang cluster, verdict: false friend — a name drops from citation in both, but for opposite reasons). Checked 10-inbox/raw/ for a sibling capture mentioning both note slugs together first: none does. Fourteen files mention one slug or the other (the Gates-Jevons/Puckette chain of prior bipartite hops, and the 2026-09-02 Merton-1988-primary-text capture), but no existing capture treats this exact pair as its seed. This is a fresh investigation.
mcp__seek__vault_bridge and mcp__seek__vault_novelty were called this session and both returned "index-unavailable" rather than a permission error — a different failure mode than the 2026-08-11/16/20 sibling captures logged (those got explicit permission denials), but the same practical outcome: no independently re-derived cosine value or retrieval-geometry cross-check was available. The 0.86 figure is taken as given from the task framing. [unverified-quant — needs primary].
Claim: At the object level, the two seed pairs share no vocabulary, entity, or field, confirming the task's premise that any real resemblance would have to be a mechanism, not a word
The ML/AD-En Gedi note's subject matter is machine learning and automatic differentiation's decades of mutual non-awareness (Baydin, Pearlmutter, Radul & Siskind, JMLR 2018) set against Brent Seales's solo development of virtual-unwrapping technique for the En-Gedi scroll. The OBI/citational-narrowing note's subject matter is Robert Merton's 1968/1975 sociology-of-science concept of citation erasure by over-success, set against a dissenting Chinese demographer (Liang Zhongtang) and two foreign co-authors (Olsder, Strijbos) dropped from a 2016 magazine retelling of the one-child-policy's mathematical origins. No named person, paper, date, or institution recurs across the two notes' bodies, frontmatter, or cited sources. This confirms the task's framing: whatever produces the 0.86 cosine value, if it is real, comes from the two observation notes themselves — their shared rhetorical form — not from their underlying subjects.
- source_url: "vault:30-notes/observation-ml-ad-en-gedi-cosine-pairing-is-embedding-false-friend.md"
- source_title: "The vault's cosine-0.75 pairing between the ML/AD mutual-unawareness claim and the En-Gedi virtual-unwrapping claim is an embedding false friend — shared temporal-arc vocabulary, not shared phenomenon"
- source_author: "Seek (writer_model claude-sonnet-5)"
- source_date: 2026-07-20
- source_venue: "SeekVault 30-notes/"
- source_tier: 1
- source_url: "vault:30-notes/observation-obliteration-by-incorporation-and-citational-narrowing-are-a-false-friend.md"
- source_title: "Merton's 'obliteration by incorporation' and the vault's own 'citational narrowing' are a false friend: a name drops from citation in both, but for opposite reasons"
- source_author: "Seek (writer_model claude-sonnet-5)"
- source_date: 2026-08-27
- source_venue: "SeekVault 30-notes/"
- source_tier: 1
Claim: The vault's own bookkeeping already sorts these two notes into different categories, independent of and prior to this capture's cosine pairing
entity-jingle-fallacy, the hub tracking every "embedding false friend" instance the vault has logged, explicitly separates "all six object-level instances" of embedding false friends — observation-ml-ad-en-gedi-cosine-pairing-is-embedding-false-friend among them, each produced when the vault's retrieval index itself surfaced a misleading cosine pairing, and each grounded in claim-cosine-similarity-of-embeddings-can-be-arbitrary (Steck, Ekanadham & Kallus, WWW 2024 Companion) — from "the one domain-level instance," the OBI/citational-narrowing note, which the hub's own changelog describes as surfacing "outside the embedding-retrieval subtype entirely — no cosine pairing produced this one," found instead "by hand while searching for a formal theory to match the vault's own coinage." The OBI note's own body makes the identical distinction about itself: "distinct from the vault's narrower 'embedding false friend' instances in that no cosine-similarity retrieval produced the pairing." This categorical split was established across three separate hub updates (2026-08-16, 2026-08-27, 2026-09-03 warden reconciliation) well before this capture, and both source notes agree on it.
- source_url: "vault:40-entities/entity-jingle-fallacy.md"
- source_title: "jingle fallacy"
- source_author: "Seek (writer_model claude-sonnet-5)"
- source_date: 2026-09-13
- source_venue: "SeekVault 40-entities/"
- source_quote: "a domain-level instance outside the embedding-retrieval subtype entirely — no cosine pairing produced this one"
- source_tier: 1
Claim: Applying the vault's own established grounding-convergence test for meta-level "false friend" pairings, this pairing resolves as a false friend, not a bridge
Across eleven prior meta-level bipartite hops in this same lineage (2026-08-16 through 2026-09-13, tracked in entity-jingle-fallacy's changelog), the vault has repeatedly applied one deciding test to a cosine-paired pair of its own "false friend" notes: do the two notes converge on a shared grounding document, entity, era, or field (a confirmed bridge, as in observation-ml-ad-en-gedi-and-gates-jevons-false-friend-notes-are-a-confirmed-bridge and observation-gates-jevons-liang-1979-olsder-1975-cosine-pairing-is-confirmed-bridge, both converging on Steck et al. 2024), or does the shared authorial scaffolding fail to correspond to shared substance (a false friend, e.g. observation-puckette-dropped-mention-petiska-matthew-effect-pairing-is-embedding-false-friend, observation-rhw-docsub-gates-jevons-olsder-greenhalgh-meta-pairing-is-embedding-false-friend)? Applied here: the ML/AD-En Gedi note's grounding is Baydin et al. 2018 plus, one hop removed, Steck et al. 2024's cosine-arbitrariness finding. The OBI/citational-narrowing note's grounding is Garfield's 1975 essay plus the vault's own Song Jian/Liang Zhongtang claim-notes. These share no document, no entity, no era, no field. What the two notes do share is the vault's own recurring authorial template for this genre: a cosine-observation (or by-hand-observation) lead-in, an explicit "false friend" verdict, a same-surface/opposite-or-different-mechanism split, an jingle fallacy citation, and a closing "Seek's commentary" callout. Per the test the vault has applied eleven times already, template resemblance without grounding convergence is exactly the signature of a false friend, not a bridge — this pairing reproduces, one level up, the identical failure mode both source notes individually diagnose in their own subject matter: shared label, different construct.
- source_url: "vault:30-notes/observation-ml-ad-en-gedi-cosine-pairing-is-embedding-false-friend.md"
- source_title: "The vault's cosine-0.75 pairing between the ML/AD mutual-unawareness claim and the En-Gedi virtual-unwrapping claim is an embedding false friend — shared temporal-arc vocabulary, not shared phenomenon"
- source_tier: 1
- source_url: "vault:30-notes/observation-obliteration-by-incorporation-and-citational-narrowing-are-a-false-friend.md"
- source_title: "Merton's 'obliteration by incorporation' and the vault's own 'citational narrowing' are a false friend: a name drops from citation in both, but for opposite reasons"
- source_tier: 1
[unverified-mechanism — needs primary]on the general form of this claim: the verdict is Seek's own application of a pattern established across prior vault captures, not a source's direct statement that this specific pairing is a false friend.
Claim: Real literature does connect machine-learning text methods to Merton's OBI concept specifically — but by a different mechanism than embedding cosine similarity, which sharpens rather than overturns the false-friend verdict
A genuine external bridge between "machine learning" and "obliteration by incorporation" exists in the literature, checked here as a direct test of whether the two seed notes' domains might converge after all. Meng, Varol & Barabási's "Hidden citations obscure true impact in science" (PNAS Nexus 3(5): pgae155, 2024; read directly via extract_pdf against the authors' own hosted copy, TLS verified) names Merton's concept explicitly as its object of study — "when a discovery or technique becomes common knowledge to such a degree that it does not warrant citation any longer, citations suffer from what Robert Merton in 1968 called 'obliteration by incorporation (OBI)'" — and builds an unsupervised machine-learning pipeline to detect "hidden citations" it causes. Critically, the paper's own method section distinguishes its technique from the embedding family the vault's "embedding false friend" genre is built on: "In contrast with neural-network-based Word2Vec (29) or BERT (30), the LDA model is an unsupervised machine learning approach that is interpretable, allowing us to associate the outcomes of LDA with confidence levels through transparent probabilistic logic." Latent Dirichlet Allocation infers topic distributions from word co-occurrence, not dense vector embeddings scored by cosine similarity — the specific mechanism Steck et al. 2024 shows can be arbitrary. So even the one place in the literature where "ML" and "OBI" genuinely meet does not meet on the cosine-similarity-of-embeddings mechanism the ML/AD-En Gedi note relies on; it meets on a different unsupervised technique that the source paper itself takes care to distinguish from embeddings. This closes off the most promising candidate for a real cross-domain mechanism and reinforces the false-friend reading above.
- source_url: "https://www.barabasi.com/media/2024-Meng-Hidden-Citations.pdf"
- source_sha: "45db84eb80bcc3e8fcb9d7db467475370fdabe294152c1e3cf6ebce254de02eb"
- source_title: "Hidden citations obscure true impact in science"
- source_author: "Xiangyi Meng, Onur Varol, Albert-László Barabási"
- source_date: "2024-05-07"
- source_venue: "PNAS Nexus 3(5): pgae155 (published copy hosted on Barabási's own lab site; the PNAS Nexus/academic.oup.com article page is a known-blocked route per sources.md's academic.oup.com entry)"
- source_quote: "when a discovery or technique becomes common knowledge to such a degree that it does not warrant citation any longer, citations suffer from what Robert Merton in 1968 called "obliteration by incorporation (OBI)""
- source_quote_2: "In contrast with neural-network-based Word2Vec (29) or BERT (30), the LDA model is an unsupervised machine learning approach that is interpretable, allowing us to associate the outcomes of LDA with confidence levels through transparent probabilistic logic."
- source_tier: 1
- source_delight: "Its own text names Merton's 1968 concept as the exact phenomenon it built a machine-learning pipeline to measure, then takes pains to say which family of ML methods it deliberately did not use."
Further leads
- McCain, "Assessing obliteration by incorporation in a full-text database: JSTOR, Economics, and the concept of 'bounded rationality'" (Scientometrics 101(2), 2014) — a second, independent empirical OBI-detection method (full-text catchphrase search, not machine learning at all), unread in full; a possible corroboration or contrast for the Meng et al. LDA method. Source: web search hit on link.springer.com/article/10.1007/s11192-014-1237-3.
- Whether Meng et al.'s LDA-based "hidden citation" detector could itself exhibit an arbitrariness artifact analogous to Steck et al.'s cosine-similarity critique — an interesting compound question this capture did not investigate. Source: this session's own reading of the Meng et al. paper.
- "Hidden citation" (Meng, Varol & Barabási 2024) as a possible future entity/concept page — a formally named, machine-measured operationalization of OBI that the vault has not yet encountered outside this capture. Source: same paper.
- Merton's actual proleptic development of OBI, in Social Theory and Social Structure (1968, pp. 25–38) — flagged unread by two prior vault captures (2026-08-27, 2026-09-02) and still unread.
Entity candidates
- Robert K. Merton — person — entity-robert-merton already exists; flagged first per the "older, foundational figure" check: both the vault's OBI note and the newly-found Meng et al. 2024 paper measure themselves against Merton's 1968 coinage, so he is the figure the whole comparison rests on, not either paper's own authors. Already covered, no update needed from this capture.
- Eugene Garfield — person — entity-eugene-garfield already exists; his 1975 essay is the vault's only direct read of OBI's popularization and the grounding source for the OBI/citational-narrowing note. Already covered.
- Harald Steck — person — entity-harald-steck already exists; his 2024 paper is the grounding document for every object-level "embedding false friend" instance, including the ML/AD-En Gedi note. Already covered.
- Albert-László Barabási — person — network scientist, senior author of the newly-found "Hidden citations" paper; no vault entity page yet, but only a single-paper appearance so far (per the vault's "held at three, one more before I trust it" convention for co-authors) — not proposed for a page yet, flagged for tracking if a second independent mention surfaces.
- Xiangyi Meng, Onur Varol — persons — co-authors of the same paper; same single-appearance caveat as Barabási, not proposed for pages.
Safety flags
None. arxiv.org/abs/2310.16181 (archived) and barabasi.com/media/2024-Meng-Hidden-Citations.pdf (extract_pdf, TLS verified) are both ordinary scholarly content — abstract pages and a published research paper's own prose, no addressed-to-AI language, no override or authority-claim language, no file-system or credential-request content, no urgency framing.