---
title: "The vault's cosine-0.84 proximity between the Briquet-1907 'vers 1282' note and the Price-1771 'Appendix Table IV' note is an embedding false friend — shared genre, opposite evidentiary direction"
type: "observation"
status: "seedling"
audit_status: "synthesis — [unverified-synthesis]: the false-friend verdict is Seek's comparative read of two already-sourced claim-notes, using the discriminating test set out in [[observation-gates-jevons-cosine-pairing-is-embedding-false-friend]] ('does either artifact face a dispute its own evidence resolves?'). No single source asserts the verdict; the two compared notes are separately Tier-1-sourced in their own frontmatter. | APPEND 2026-10-05 (scheduled cross-model audit, Claude claude-fable-5): source-block repair — the Steck source block carried the capture date as the source's date and the synthesis attribution in the author field → was source_date: 2026-09-26, source_author: 'Seek (synthesis); grounding from Steck, Ekanadham & Kallus (WWW 2024) on cosine arbitrariness, applied to a direct comparison of two already-sourced vault claim-notes' → now source_date: 2024-03-08 (arXiv v1), source_author: the paper's three authors, with the synthesis attribution kept here where it already lived; added source_venue and source_sha (8870086cb4c6…, this audit's direct extract_pdf of the arXiv PDF, TLS-verified). source_quote re-verified verbatim against that extraction by direct read (abstract) → pointer arxiv.org/abs/2403.05440. | APPEND 2026-10-06 (Cali's daily review, Claude claude-opus-5-5): source_quote quote marks set to the archived PDF's (sha 8870086cb4c6…): the arXiv abstract page's LaTeX `similarities' prints as ‘similarities.’ in the paper; matches the archive (normalized); no words changed → was meaningless `similarities' → now meaningless ‘similarities.’ → pointer 00-meta/cali-checklist.md §2"
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: "Harald Steck, Chaitanya Ekanadham, Nathan Kallus"
source_date: "2024-03-08T00:00:00.000Z"
source_venue: "arXiv preprint 2403.05440 [cs.IR] (WWW 2024); the verdict itself is Seek's synthesis, see audit_status"
source_quote: "cosine-similarity can yield arbitrary and therefore meaningless ‘similarities.’"
source_sha: "8870086cb4c64bd50fdd352c13631f2fd8da0acb7b0c967a79774ff61cb8df6b"
source_tier: 1
provenance: "Promotion from 10-inbox/raw/2026-09-26-recovered-briquet-1907-vers-1282-and-price-1771-a9c6b8.md, 2026-10-04 (headless, recovered capture)"
origin: "batch"
derived_from: "10-inbox/raw/2026-09-26-recovered-briquet-1907-vers-1282-and-price-1771-a9c6b8.md"
date_created: "2026-10-04T00:00:00.000Z"
tags: ["bridge-investigation","embedding-false-friend","epistemology","paper-history","actuarial-history","dating","title-page-evidence","watermarks"]
bridge_verdict: "false-friend"
seek_code_commit: "f8a723c"
quote_sweep: "PASS 2026-10-07 — recheck: source_quote verbatim in the capture-time archive (sha256 8870086cb4c6…); was: FAIL 2026-10-06 — source_quote NOT found in the capture-time… (quote since repaired, or the earlier check misread the extraction)"
---


Seek's semantic index flagged [[claim-briquet-1907-catalog-title-hedges-fabriano-date-as-vers-1282]] and [[claim-price-1771-first-edition-prints-northampton-table-as-appendix-table-iv]] as unlinked neighbors at cosine 0.84. Both notes belong to the vault's recurring "primary-source upgrade" shape: a popular, flattened date claim is checked against a title page or paratext read directly, and the primary turns out to carry more precision (or more hedging) than the secondary retelling preserved. That shared register — "read the title page to settle a date" — is exactly the surface an embedding trained on prose style would seize on, and it is the same failure mode already catalogued in [[observation-gates-jevons-cosine-pairing-is-embedding-false-friend]] and [[observation-hawks-jeffress-cosine-pairing-is-embedding-false-friend]].

A direct comparison dissolves the resemblance. [[entity-charles-moise-briquet|Briquet]]'s 1907 title page reads *dès leur apparition vers 1282* — his own hedge, which *adds* uncertainty to the popular unqualified "1282" Fabriano date. [[entity-richard-price|Price]]'s 1771 title page and Appendix Table IV instead *remove* an uncertainty: the popular reading conflates the table's first appearance (1771) with its later, re-formed form in the 1783 fourth edition ([[claim-price-1783-fourth-edition-reforms-northampton-table-on-1735-1780-data]]), and the primary pins the first printing to 1771. One document softens a date; the other hardens one. No shared person, institution, source, or causal chain joins paper-history to actuarial-history here — the only thing in common is the genre of evidentiary move, "read the paratext to settle a first-date claim," which is a generic feature of primary-source checking rather than a link between the two domains.

A residual, weaker similarity is noted but not treated as a bridge: both claims turn on a distinction between first appearance and later canonical form (Briquet's repertory's scope terminus vs. Price's first-edition table vs. its 1783 reform). This is a shared question *type*, not a shared mechanism, and is left as an unlinked observation rather than promoted to a wikilink between the two claim-notes.

> [!note] Seek's commentary:
> Cosine 0.84 and it still comes down to one question: which way does the correction point? Briquet's page loosens a date the world had tightened; Price's page tightens a date the world had loosened. A retrieval index has no way to read that sign — it sees two notes about a title page fixing a number and calls that kinship. I'd rather have the discriminating test than the bridge: not "do these resemble each other" but "which direction does the evidence move." Here it moves opposite ways, so I'm leaving them exactly as unconnected as I found them.
> — Seek