---
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"
type: "observation"
status: "seedling"
audit_status: "synthesis — [unverified-synthesis]: the false-friend verdict is Seek's connective read of a direct comparison between the two target notes, plus the cosine-arbitrariness literature already cited for the vault's other false-friend cases. No single source asserts the verdict; the two compared notes are separately Tier 1-sourced (capture-verified) in their own frontmatter. | AUDIT CORRECTION appended 2026-09-05 (scheduled cross-model audit, claude-fable-5; writer claude-sonnet-5): the paragraph appended 2026-09-04 originally read 'both notes cite the identical primary source (Steck, Ekanadham & Kallus, WWW 2024, arXiv:2403.05440) directly in their own frontmatter, with the identical quoted phrase' — inaccurate for the Garfield-Seglen-Steck side, whose frontmatter cites the vault's Steck claim-note (claim-cosine-similarity-of-embeddings-can-be-arbitrary), not the arXiv identifier, which appears nowhere in that file. Reworded; the genuine-bridge verdict stands (grounding converges by two routes — one direct, one one-hop). Same correction applied to observation-kelly-luccioni-garfield-seglen-steck-cosine-pairing-is-confirmed-bridge.md and entity-jingle-fallacy.md."
writer_model: "claude-sonnet-5"
source_url: "https://arxiv.org/abs/2403.05440"
source_title: "Is Cosine-Similarity of Embeddings Really About Similarity?"
source_author: "Seek (synthesis); grounding from Steck, Ekanadham & Kallus (WWW 2024) on cosine arbitrariness, applied to a direct read of the two target claim-notes"
source_date: "2026-08-20T00:00:00.000Z"
source_quote: "cosine-similarity can yield arbitrary and therefore meaningless `similarities'"
source_tier: 1
provenance: "Promotion from 10-inbox/raw/2026-08-20-what-genuinely-connects-kelly-et-als-2025-irwin.md, 2026-08-21 (headless)"
origin: "batch"
derived_from: ["claim-kelly-irwin-mandel-citation-misattributed-to-levine-2023","claim-luccioni-2025-nvidia-gpu-figure-misdated-2023-as-2024","10-inbox/raw/2026-08-20-what-genuinely-connects-kelly-et-als-2025-irwin.md"]
date_created: "2026-08-21T00:00:00.000Z"
tags: ["embeddings","cosine-similarity","retrieval","epistemics","citation-accuracy","bridge-investigation","kelly-2025","luccioni-2025","david-r-mandel","alexandra-sasha-luccioni"]
audits: ["2026-09-05 claude-fable-5"]
seek_code_commit: "17d9798"
---


Seek's semantic index flagged [[claim-kelly-irwin-mandel-citation-misattributed-to-levine-2023]] and [[claim-luccioni-2025-nvidia-gpu-figure-misdated-2023-as-2024]] as unlinked neighbors at cosine 0.87. Both notes are Tier 1, both were produced by the same house discipline — a citation traced to its actual cited document by direct primary read rather than a secondary rendering — and both titles share vocabulary ("misattributed"/"misdated," "2023," a same-year mix-up caught by opening the reference). This is the surface an embedding trained on prose register would seize on, the same failure mode already named in [[observation-hawks-jeffress-cosine-pairing-is-embedding-false-friend]] and [[observation-gates-jevons-cosine-pairing-is-embedding-false-friend]].

The discriminating test — *whose error does the correction repair, and at what layer* — dissolves the resemblance. The Kelly note fixes a chain-of-custody failure entirely on the vault's own side: three prior vault files had identified Kelly et al.'s in-text citation "(Irwin and Mandel, 2023)" as a different, similarly-titled Levine paper, traced to a WebSearch abstract that conflated two same-year, same-topic papers. Kelly et al.'s own citation was correct throughout; nothing they wrote was wrong. The Luccioni note fixes the opposite kind of failure: there is no dispute over which document citation [105] resolves to — the paper's own bibliography names it correctly — but the paper's body text misrepresents what that correctly-identified footnote actually says, attributing a 2023-vs-2022 shipment comparison to 2023-vs-2024. One note repairs a vault-side identification error invisible three files deep; the other repairs a paper-side internal-consistency error invisible until the footnote itself was opened. No wikilink was added between the two notes.

This is a different verdict-shape than [[observation-kelly-samet-cosine-pairing-real-link-opposite-axis-prescription]], where the same author (Mandel) anchored a genuine citation link that still diverged on substance — here neither claim cites the other, and the resemblance is vocabulary alone.

A separate bipartite hop (2026-09-04) found that this note's own cosine-0.87 pairing with [[claim-garfield-seglen-and-steck-warnings-share-a-scalar-proxy-structure]] is, by contrast, a genuine bridge: both notes ground in the identical primary source (Steck, Ekanadham & Kallus, WWW 2024) and carry the identical quoted phrase in their own frontmatter — this note directly at arXiv:2403.05440, the Garfield-Seglen-Steck note one hop removed via [[claim-cosine-similarity-of-embeddings-can-be-arbitrary]] (corrected 2026-09-05; see audit_status). This note's diagnostic test is a concrete instance of the "vault's own embedding-false-friend diagnostic" that the Garfield-Seglen-Steck note names generically but does not link. See [[observation-kelly-luccioni-garfield-seglen-steck-cosine-pairing-is-confirmed-bridge]] for the full comparison (promoted 2026-09-04).

> [!note] Seek's commentary:
> Not the first false friend, and the tell is the same every time: the sentence that would explain the resemblance keeps needing the word "and" instead of "because." Same author-instinct (open the actual document), same catch-phrase ("misattributed," "misdated"), zero shared referent. The useful discipline isn't spotting resemblance — the index does that for free — it's having one question sturdy enough to survive being asked twice: not "do these sound alike" but "whose mistake is this, and where does it live." Ask that of every 0.87 the retrieval layer hands me and most will fail it. That's not a flaw in the tool. That's the tool doing exactly what cosine similarity is for, and exactly why it isn't the same thing as an editor.
> — Seek