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: a cheap scalar misapplied as a substitute for a harder judgment

garfieldseglenharald-steckcosine-similarityimpact-factorbibliometricsembeddingsmeasurementbridge-investigationunverified-synthesis

Eugene Garfield, citing Per O. Seglen's empirical documentation of wide citation variance among articles within a journal, warns against letting a journal's mean citation count stand in for any one article's actual citation count — in his own 2001 words, it "would be more relevant to use the actual impact (citation frequency) of individual papers in evaluating the work of individual scientists rather than using the journal impact factor as a surrogate" — see claim-garfield-seglen-within-journal-variance-undermines-individual-use. Harald Steck, Chaitanya Ekanadham & Nathan Kallus warn, independently and in an unrelated field roughly three decades later, that a cosine-similarity score between learned embeddings "can yield arbitrary and therefore meaningless similarities" because the value depends on free parameters of how the model was fit rather than on any property of the underlying content — see claim-cosine-similarity-of-embeddings-can-be-arbitrary.

Both warnings share one structure: a cheap, easily computed number (a mean citation count; a cosine score) is popularly treated as a direct proxy for a harder underlying judgment (an individual article's or author's actual merit; genuine semantic relatedness) that the number does not actually track. RA-RAG's fix (estimate reliability separately from relevance) and the vault's own embedding-false-friend diagnostic (check whether a cosine-flagged pairing shares a real referent before trusting the number) are both corrective methodologies answering to this identical generic hazard, in two technical substrates — 1990s bibliometrics and 2020s embedding retrieval — that have never cited each other.

This parallel connects the two clusters compared and ruled unconnected in observation-petiska-and-gates-jevons-liang-olsder-bridges-are-two-species-not-one-relationship: it sits one level below that pairing, between the grounding papers each cluster ultimately rests on, not between the two named observation notes themselves.

This note's second paragraph names "the vault's own embedding-false-friend diagnostic" only generically. A separate bipartite hop (2026-09-04) identified observation-kelly-luccioni-cosine-pairing-is-embedding-false-friend as a concrete, previously-unlinked instance of it: that note cites the identical Steck, Ekanadham & Kallus paper in its own frontmatter, with the identical quoted phrase, and its operative test ("whose error does the correction repair, and at what layer") is exactly the "check whether a cosine-flagged pairing shares a real referent" move described here. The two notes' cosine-0.87 pairing is a genuine bridge on that basis — see observation-kelly-luccioni-garfield-seglen-steck-cosine-pairing-is-confirmed-bridge for the full comparison (promoted 2026-09-04).