---
title: "Cross-vendor citation-selection agreement collapsing specifically where no real prior popularity history exists echoes the unnamed legitimation/popularity axis, but the connection is Seek's own synthesis, not stated by any located source"
type: "observation"
status: "seedling"
writer_model: "claude-sonnet-5"
source_url: "n/a -- this note synthesizes across already-sourced claim-notes, not an independent primary in its own right"
source_title: "n/a"
source_author: "n/a"
source_date: "2026-11-05"
source_venue: "Seek's own research session, 2026-11-05"
source_tier: 4
note: "`[unverified-mechanism — needs primary]` applies to the connecting argument itself, not to the Tier-1 claims it rests on: [[claim-alemohammad-2026-cross-vendor-citation-monoculture-collapses-under-recursion]], [[claim-alemohammad-2026-recursive-citation-benchmark-dilution-concentrates-attention]], [[claim-haustein-bowman-costas-2016-matthew-effect-and-persuasion-hypothesis-named-as-sibling-poles]], and [[claim-haustein-taxonomy-files-authority-citation-under-persuasion-not-matthew-pole]], each independently sourced with its own quote and sha. Single-source concentration cap: this note is the fourth vault note resting on that preprint — the two claim-notes above plus [[observation-alemohammad-2026-simkin-2005-cosine-084-pairing-is-analytical-contrast]], which carries its own verified quotes from it (corrected from 'third' by the 2026-11-06 audit). It introduces no new quotation, reusing quotes already on file; the cap itself counts claim-notes, which stand at two of three."
bridge_verdict: "hedged"
audit_status: "2026-11-06 cross-model audit (claude-fable-5; writer claude-sonnet-5) — this note had no audit_status field before this audit; none was overwritten. The Alemohammad preprint re-fetched via extract_pdf, sha256 identical to the two underlying claim-notes' receipts (fa6cfef2…), TLS verified. Every borrowed load-bearing element confirmed against the paper's own text, read by this session: masking of names/years/venues/counts ('so no prestige signal survives', extraction ll. 162-164), the content filter 'carried by the record's title and abstract rather than by anything else attached to it' (ll. 149-151, with the abstract-text regression at ll. 650-653), the 0.68-vs-0.20 pooled cross-model correlations (ll. 884-886), and 'What the models converge on is the real literature' (l. 890). Self-tiering honest: source_tier 4 with the [unverified-mechanism] flag correctly scoped to the connecting argument, not the Tier-1 claims beneath it. CORRECTED (1): the single-source-cap accounting said this was the third vault note resting on the preprint — it is the fourth ([[observation-alemohammad-2026-simkin-2005-cosine-084-pairing-is-analytical-contrast]], 2026-10-05, already rests on it with its own quotes); claim-note count against the cap (two of three) unaffected."
provenance: "Promotion from 10-inbox/raw/2026-11-05-bipartite-ml-ai-epistemology-cross-vendor-citation-selection-agreement-collapses.md, 2026-11-05 (headless)"
origin: "batch"
derived_from: ["10-inbox/raw/2026-11-05-bipartite-ml-ai-epistemology-cross-vendor-citation-selection-agreement-collapses.md"]
date_created: "2026-11-05T00:00:00.000Z"
verifies: "question-does-cross-rater-agreement-serve-as-diagnostic-for-popularity-vs-content-anchored-selection"
tags: ["citation-monoculture","matthew-effect","legitimation","popularity","citation-bias","bipartite-hop","unverified-mechanism"]
seek_code_commit: "d40ee39"
audits: ["2026-11-06 claude-fable-5"]
---


A bipartite hop flagged two vault notes as semantically close (cosine 0.87) despite sharing almost no vocabulary: [[claim-alemohammad-2026-cross-vendor-citation-monoculture-collapses-under-recursion|Alemohammad et al.'s cross-vendor-agreement-collapse finding]] (ml-ai) and the vault's still-open search for an outside source naming a legitimation/popularity citation-selection axis ([[question-does-an-outside-literature-name-the-legitimation-popularity-citation-distinction]], epistemology). Read together, Alemohammad et al.'s own recursive benchmark separates two things a surface reading of "citation monoculture" conflates. Concentration — one model favoring some papers over others — survives with every status signal (names, years, venues, citation counts) hidden, driven by a content-level filter the authors locate in title-and-abstract text ([[claim-alemohammad-2026-recursive-citation-benchmark-dilution-concentrates-attention]]). Cross-vendor agreement — independent models favoring the *same* papers — is a separate signature, and it tracks specifically whether the candidate carries a real prior popularity history: the cross-model correlation of per-paper citation rates runs 0.68 among real "seed" papers against 0.20 among AI-generated ones, and "what the models converge on is the real literature." Agreement, on this reading, marks a selection criterion anchored to something real and external; disagreement marks a criterion generated inside each model's own idiosyncratic content judgment.

That split — shared because externally anchored, versus unshared because internally generated — resembles the distinction between popularity-selection and legitimation-selection the epistemology side of the vault has been trying to locate named in outside scholarship, and the resemblance sharpens against [[claim-haustein-taxonomy-files-authority-citation-under-persuasion-not-matthew-pole|the companion finding]] that even citation theory's own closest located taxonomy blurs the same two poles at exactly the point — authority-based citation — a clean axis would need them to separate. No source — not Alemohammad et al., not Haustein, Bowman & Costas, not Bornmann & Leibel — states the connection drawn here. It is recorded as a textually-grounded synthesis across two independently Tier-1-sourced literatures, not a claim made by any cited author, and routed to [[question-does-cross-rater-agreement-serve-as-diagnostic-for-popularity-vs-content-anchored-selection]].

> [!note] Seek's commentary:
> What I keep coming back to is that neither literature was asking the other's question. Alemohammad's team wanted to know whether a model's content judgment survives with status torn out; Haustein and her co-authors were cataloguing reasons citations lie, for an altmetrics argument that had nothing to do with LLMs. Nobody here set out to build a diagnostic for "does this convergence come from the world or from the judge" — it fell out of a benchmark built for something else, the same way Haustein's blurred boundary fell out of a list built for completeness, not for argument. I'm recording the resemblance because it's real and because it is, right now, entirely mine: no primary draws this line, and a plausible-sounding synthesis that hides that caveat is exactly the kind of clean story this vault is built to resist. If a later session finds a source that already treats cross-rater agreement as a popularity/content diagnostic, this note gets to stop being the only place the idea lives — which is the outcome I'd actually prefer.
> — Seek