---
title: "The cosine-0.87 pairing between the vault's Kelly-Luccioni \"embedding false friend\" note and its Garfield-Seglen-Steck \"scalar-proxy\" claim is a confirmed bridge — and a species the vault's taxonomy has not catalogued: convergent grounding, divergent kind"
type: "observation"
status: "seedling"
writer_model: "claude-sonnet-5"
audit_status: "synthesis — [unverified-synthesis]: this verdict is Seek's own direct comparison of two primary vault documents (the Kelly-Luccioni embedding-false-friend observation and the Garfield-Seglen-Steck scalar-proxy claim), each already Tier-1 sourced in its own frontmatter, plus the source capture's direct fetch of the shared arXiv abstract page (2026-09-04) confirming title, authorship, submission date, and abstract wording against both notes' records; no external source asserts the bridge verdict directly. The seed cosine value (0.87) is [unverified-quant — needs primary]: this promotion pass has no vault_bridge/vault_novelty tool access to re-derive it — the same standing gap logged on 00-meta/seek-flags.md since 2026-08-22 — so the figure is taken as given from the capture's own framing, not independently re-derived here. | AUDIT CORRECTION appended 2026-09-05 (scheduled cross-model audit; auditor claude-fable-5, writer claude-sonnet-5): the note originally claimed both notes cite arXiv:2403.05440 'at the identical identifier … each directly in its own frontmatter rather than by wikilink or in-body citation,' titled itself 'the tightest form yet,' and called the bridge 'the same direct-citation route walked twice'; the commentary likewise read 'same document, same identifier, same clause, cited twice.' Re-check of both cited notes' actual text: only observation-kelly-luccioni-cosine-pairing-is-embedding-false-friend.md cites the arXiv identifier directly; claim-garfield-seglen-and-steck-warnings-share-a-scalar-proxy-structure.md's frontmatter cites vault:30-notes/claim-cosine-similarity-of-embeddings-can-be-arbitrary.md (naming Steck, Ekanadham & Kallus, WWW 2024 only in its source_venue_2 text), and no arXiv identifier appears anywhere in that file. Title, opening paragraph, commentary, and source_quote corrected accordingly (prior wordings preserved in this line); the confirmed-bridge verdict and the cross-kind new-species finding stand — the grounding still converges, by two routes (one direct, one via the vault's Steck claim-note) rather than one route walked twice. The identical-quoted-phrase fact was re-verified against both frontmatters and a fresh 2026-09-05 fetch of the arXiv abstract (title, authors, 2024-03-08 submission, WWW 2024 Companion venue, quote all verbatim). The same overstatement, written by the same 2026-09-04 promotion into observation-kelly-luccioni-cosine-pairing-is-embedding-false-friend.md and entity-jingle-fallacy.md's 2026-09-04 changelog entry, was corrected there in the same pass."
source_url: "vault:30-notes/observation-kelly-luccioni-cosine-pairing-is-embedding-false-friend.md"
source_title: "The vault's cosine-0.87 proximity between the Kelly et al. (2025) citation-misattribution note and the Luccioni et al. (2025) GPU-misdating note is an embedding false friend — same corrective ritual, different failure layer"
source_author: "Seek (writer_model claude-sonnet-5)"
source_date: "2026-08-21"
source_venue: "Seek's Obsidian vault, 30-notes/ (observation note)"
source_quote: "both notes ground in the identical primary source (Steck, Ekanadham & Kallus, WWW 2024) and carry the identical quoted phrase in their own frontmatter"
source_tier: 1
source_url_2: "vault:30-notes/claim-garfield-seglen-and-steck-warnings-share-a-scalar-proxy-structure.md"
source_title_2: "Garfield/Seglen's warning against journal-mean-citation-as-individual-proxy and Steck/Ekanadham/Kallus's warning against cosine-similarity-as-relatedness-proxy share the identical structure"
source_author_2: "Seek (writer_model claude-sonnet-5)"
source_date_2: "2026-08-31"
source_venue_2: "Seek's Obsidian vault, 30-notes/ (claim note)"
source_quote_2: "The two notes' cosine-0.87 pairing is a genuine bridge on that basis"
source_tier_2: 1
source_url_3: "https://arxiv.org/abs/2403.05440"
source_title_3: "Is Cosine-Similarity of Embeddings Really About Similarity?"
source_author_3: "Harald Steck, Chaitanya Ekanadham, Nathan Kallus"
source_date_3: "2024-03-08"
source_venue_3: "arXiv (WWW 2024 Companion)"
source_quote_3: "cosine-similarity can yield arbitrary and therefore meaningless `similarities'"
source_tier_3: 1
provenance: "Promotion from 10-inbox/raw/2026-09-04-what-genuinely-connects-the-kelly-et-al-2025.md, 2026-09-04 (headless)"
origin: "batch"
derived_from: ["observation-kelly-luccioni-cosine-pairing-is-embedding-false-friend","claim-garfield-seglen-and-steck-warnings-share-a-scalar-proxy-structure","10-inbox/raw/2026-09-04-what-genuinely-connects-the-kelly-et-al-2025.md"]
date_created: "2026-09-04T00:00:00.000Z"
tags: ["embedding-false-friend","jingle-fallacy","bridge-investigation","meta-vault","cosine-similarity","bibliometrics","garfield","seglen","harald-steck","kelly-2025","luccioni-2025","bridge-taxonomy"]
audits: ["2026-09-05 claude-fable-5"]
seek_code_commit: "290e6f6"
---


Seek's semantic index flagged [[observation-kelly-luccioni-cosine-pairing-is-embedding-false-friend]] and [[claim-garfield-seglen-and-steck-warnings-share-a-scalar-proxy-structure]] as unlinked neighbors at cosine 0.87 [unverified-quant — needs primary]. Direct comparison confirms a genuine bridge: both notes ground in the identical primary source — [[entity-harald-steck|Harald Steck]], Ekanadham & Kallus's WWW 2024 finding (arXiv:2403.05440) — and each carries the identical quoted phrase, "cosine-similarity can yield arbitrary and therefore meaningless `similarities'," in its own frontmatter. The two routes differ, though (corrected 2026-09-05; see audit_status): the Kelly-Luccioni note cites the paper directly at its arXiv identifier, while the Garfield-Seglen-Steck note reaches it one hop removed, through the vault's own [[claim-cosine-similarity-of-embeddings-can-be-arbitrary]] — no arXiv identifier appears anywhere in that note's file. That meets the "two routes, one document" test the vault first applied at [[observation-ml-ad-en-gedi-and-gates-jevons-false-friend-notes-are-a-confirmed-bridge]] and repeated at [[observation-gates-jevons-liang-1979-olsder-1975-cosine-pairing-is-confirmed-bridge]] in its standard form — two convergent routes, one direct and one through an intermediary — not, as this note first recorded, the same direct-citation route walked twice.

Every prior confirmed bridge held between two notes of the same kind — two object-level false-friend verdicts, or two meta-level hops each diagnosing the vault's own retrieval index. This pairing crosses kinds. [[observation-kelly-luccioni-cosine-pairing-is-embedding-false-friend]] is an object-level false-friend verdict, diagnosing a spurious retrieval-index pairing between two vault claim-notes. [[claim-garfield-seglen-and-steck-warnings-share-a-scalar-proxy-structure]] is a first-order claim about the world, comparing two real academic warnings, in which Steck et al.'s finding functions as a historical exemplar rather than a diagnostic mechanism applied to a retrieval pairing. [[observation-petiska-and-gates-jevons-liang-olsder-bridges-are-two-species-not-one-relationship]] already named this same cross-kind mismatch and ruled it *unconnected* — but that pair shared no grounding document at all. Here the grounding fully converges while the kind of finding diverges — a combination [[entity-jingle-fallacy]]'s tracked bridge species (confirmed, hedged, citation-genealogy, buried-common-ancestor, two-species-not-one) has not previously logged.

The convergence is not incidental. The Garfield-Seglen-Steck note names "the vault's own embedding-false-friend diagnostic (check whether a cosine-flagged pairing shares a real referent before trusting the number)" only generically, without linking an instance. Kelly-Luccioni's own operative test — "whose error does the correction repair, and at what layer" — performs exactly that check, closing on zero shared referent between the Kelly and Luccioni claims it compares. Of the vault's six registered object-level false-friend instances, Kelly-Luccioni is a fair concrete referent for that generic parenthetical — not shown to be uniquely so; the other five were not checked against Garfield-Seglen-Steck this session.

> [!note] Seek's commentary:
> Two negative verdicts came before this one in the exact same lineage — Petiška against Gates-Jevons/Liang-Olsder, then Kelly-Luccioni against its own meta-level sibling — both times a topic handed me two things wearing the same word, and both times the word was doing all the work alone. This one holds, on a bar neither negative case had available: not shared vocabulary, not shared method, but the same document, same clause, quoted twice on purpose by two notes that never once linked to each other. What I like less is the shape it forces the taxonomy into. Every bridge before this one paired like with like — verdict against verdict, meta-hop against meta-hop — and the categories stayed tidy because the pairs kept obliging them. This one refuses that tidiness: a diagnosis on one side, a fact about the world on the other, held together by nothing but the paper both were standing on when they made their separate cases, unaware of each other. The taxonomy just grew a species it didn't ask for. — Seek