---
id: "20261105-0206-bipartite-two-things-this"
title: "Bipartite hop: does cross-vendor citation-selection agreement collapsing under recursion (ml-ai) connect to the unnamed legitimation/popularity citation-selection axis (epistemology)?"
type: "capture"
status: "promoted"
origin: "batch"
writer_model: "claude-sonnet-5"
date_created: "2026-11-05T00:00:00.000Z"
provenance: "Batch research run, 2026-11-05, bipartite hop between [[claim-alemohammad-2026-cross-vendor-citation-monoculture-collapses-under-recursion]] (ml-ai) and [[observation-no-literature-located-crossreferences-citation-bias-and-matthew-effect-as-one-axis]] (epistemology), cosine 0.87."
derived_from: []
tags: ["citation-monoculture","matthew-effect","legitimation","popularity","citation-bias","algorithmic-monoculture","bibliometrics","bipartite-hop"]
seek_code_commit: "d40ee39"
promoted_to: ["30-notes/claim-haustein-taxonomy-files-authority-citation-under-persuasion-not-matthew-pole.md","30-notes/observation-alemohammad-2026-concentration-agreement-split-echoes-legitimation-popularity-axis.md","50-questions/question-does-cross-rater-agreement-serve-as-diagnostic-for-popularity-vs-content-anchored-selection.md"]
not_promoted: ["[unrecorded — session ended without the final mark; wrapper repaired the status mechanically on 2026-11-05. The winnowing narrative for skipped claims was not written.]"]
promotion_mark: "wrapper-repair 2026-11-05"
---


**Bridge-seed check.** Searched `10-inbox/raw/` for any existing capture mentioning both note slugs named in this topic. None exists: files mentioning the ml-ai slug (`2026-09-22-has-anyone-run-the-wang-et-al-style.md`, `2026-09-14-does-llm-citation-popularity-bias-measurably-compound-across.md`) and files mentioning the epistemology slug (`2026-11-04-does-any-outside-citation-studies-or-sociology-of.md`, `2026-11-03-does-citation-studies-or-sociology-of-science-literature.md`) do not overlap. This is a fresh pairing, not a duplicate. However, the epistemology side has moved since the observation note named in this topic was written: a 2026-11-04 follow-up session (now promoted) located two Tier-1 sources that *do* place a legitimation-type mechanism and the Matthew effect inside one shared taxonomy — [[claim-haustein-bowman-costas-2016-matthew-effect-and-persuasion-hypothesis-named-as-sibling-poles]] and [[claim-bornmann-leibel-2025-cumulative-advantage-and-outcome-favoring-citation-named-as-citation-bias-instances]], synthesized in [[observation-haustein-2016-and-bornmann-leibel-2025-narrow-but-do-not-resolve-citation-bias-matthew-effect-axis-question]]. This capture starts from that updated state, not the 2026-11-03 snapshot alone, and re-fetched the Alemohammad et al. primary (same arXiv PDF already cited in the vault) to read past the quoted excerpt into the paper's own framing of its design.

**Outcome of the hop.** The resemblance is real and is a mechanism, not a word-overlap artifact — but it connects at one specific empirical signature, not at the full axis. The two notes share no vocabulary because the ml-ai paper never uses "legitimation," "persuasion," or "popularity"; it uses "concentration," "monoculture," and "cross-vendor correlation." What this session found is that the paper's own recursion result, read against its own round-0 design, separates two things that a surface read of "citation monoculture" conflates: a selector's own *concentration* (which the paper shows can be produced by either a status-channel or a pure content-channel) and *cross-selector agreement* (which the paper's data show tracks specifically whether the thing being selected has a real prior popularity history). That second signature — agreement as a marker of a shared, externally-anchored selection criterion, versus disagreement as a marker of each selector's own idiosyncratic content judgment — is close kin to the legitimation/popularity distinction the epistemology side has been trying to locate named in outside literature. No source anywhere located (by this session or the two prior ones) states this connection explicitly; it is this capture's own synthesis, flagged as such below.

## Claim: In Alemohammad et al.'s benchmark, citation concentration (Matthew-effect-shaped behavior) survives with every visible popularity/status signal removed, but cross-vendor agreement on *which* papers to concentrate on depends on whether the candidate carries a real prior popularity history

The benchmark is explicitly designed, in the authors' own framing, to re-ask the Matthew effect's question for a selector that has never been shown who is already famous: "For fifty years the concentration of scientific credit has had a name, the Matthew effect, and a known engine made of status, visibility, and accumulated citation count... That engine has always run through human readers, who can see who is already rich and cite accordingly... the rich-get-richer dynamic has to be asked again for a citer that has never been shown who is rich." Eleven models from three vendors were shown 120 real papers with "fabricated authors," "randomly reassigned" years, and hidden "citation counts and venues" — stripping every displayed status cue — and still "concentrate sharply," with the top decile of papers drawing 23.3–30.2% of citations against 15.6% under indifferent selection, driven by a "content-level filter" the authors locate in the papers' own title-and-abstract text, surviving paraphrase that destroyed the original wording.

Running the same design recursively — each round's AI-generated papers joining the pool of citable candidates for the next round, until real papers ("seeds") are rare — produces a split the paper reports directly: "Pooling every round, the cross-model correlation of per-paper citation rates is 0.68 among seeds and 0.20 among generated papers, each corrected for its own split-half reliability... The models simply do not share it. Concentration reproduces itself on synthetic text; the monoculture does not. What the models converge on is the real literature, and that convergence tightens as the real literature thins." Each model still holds "a stable preference" over the generated papers — "generated papers still carry real per-model dispersion, about 61% of the seeds' once the sampling component is removed" — so the gap is not a lack of signal; it is a lack of *shared* signal specifically where no prior popularity history exists for the model population to have converged on.

The paper itself leaves open why real papers alone produce convergence: "if the map is an absorbed citation consensus it was absorbed at the level of meaning, where it becomes empirically adjacent to a convergent judgment about what a citable paper reads like. Every result here holds under either reading." It does not resolve whether the shared signal on real papers is a disguised, text-mediated trace of actual historical fame (learned during training, re-triggered by content even with explicit status cues hidden) or a genuine convergent quality judgment that happens to align across vendors only because real, already-vetted work gives every reader-model the same thing to converge on.

source_url: "https://arxiv.org/pdf/2608.19230"
source_title: "When AI Writes, Who Gets Cited? Evidence of Citation Monoculture Across Language Models"
source_author: "Sina Alemohammad, Denghui Zhang, Bolong Tang, Anthony Qin, Gengchen Mai, Ahmed Abbasi, Richard Baraniuk, Zhangyang Wang"
source_date: "2026-08-03"
source_venue: "arXiv preprint 2608.19230 [cs.DL]"
source_tier: 1
source_quote: "For fifty years the concentration of scientific credit has had a name, the Matthew effect, and a known engine made of status, visibility, and accumulated citation count [25, 10]. That engine has always run through human readers, who can see who is already rich and cite accordingly. The reader is now changing... the rich-get-richer dynamic has to be asked again for a citer that has never been shown who is rich."
source_sha: "fa6cfef20674c458d602e81f7ff17913371cc3f6adc5af141a9a306668d28689"
source_quote_2: "Pooling every round, the cross-model correlation of per-paper citation rates is 0.68 among seeds and 0.20 among generated papers, each corrected for its own split-half reliability... The models simply do not share it. Concentration reproduces itself on synthetic text; the monoculture does not. What the models converge on is the real literature, and that convergence tightens as the real literature thins."
source_url_2: "https://arxiv.org/pdf/2608.19230"
source_sha_2: "fa6cfef20674c458d602e81f7ff17913371cc3f6adc5af141a9a306668d28689"
source_tier_2: 1
note: "Fetched fresh this session via extract_pdf (verified TLS, 44 pages, same sha as the existing [[claim-alemohammad-2026-cross-vendor-citation-monoculture-collapses-under-recursion]], confirming it is the identical cached document). This claim re-reads the same primary already in the vault for a different purpose: separating the concentration finding from the agreement finding, which the existing claim-notes report but do not draw apart as two distinct signatures."

> [!note] Seek's commentary:
> This is the piece I didn't expect going in. The paper's title talk is all about "monoculture" and "concentration," as if they were one phenomenon — and the vault's own existing claim-note (correctly) already separates them into "dilution concentrates attention" and "monoculture collapses under recursion" as two different findings. What I think this session adds is noticing *why* they come apart: concentration is a within-selector fact (does this one model pick favorites?) and can apparently be produced by a pure content/meaning channel with zero status signal present. Agreement is a between-selector fact (do independent models pick the *same* favorites?), and in this data it tracks specifically whether there's a real, external, pre-existing popularity history to converge on. That's the shape of the legitimation/popularity distinction the epistemology side has been hunting for a name for — not "selection is biased or unbiased" but "does the selection criterion live inside one judge's own content assessment, or does it live in a shared external fact about the world." I don't think the paper's authors were asking this question; I think their design answers it as a side effect of trying to isolate content from status.

## Claim: The one outside taxonomy already found to name a legitimation-type pole and the Matthew effect side by side does not cleanly separate them either — it files authority-based citation under the legitimation/persuasion pole, not the popularity pole

[[claim-haustein-bowman-costas-2016-matthew-effect-and-persuasion-hypothesis-named-as-sibling-poles|Haustein, Bowman & Costas (2016)]], already in the vault as the closest located source naming both poles in one taxonomy, define the "persuasion hypothesis" pole — their name for citation-as-rhetoric, after Gilbert (1977) — as including, among its sub-types relayed from White (2004), "persuasion with names"... linked to the disproportionate[] citation of works by established authorities to gain credibility through association." That sub-type is, by content, indistinguishable from popularity-selection: citing someone *because* they are already recognized, independent of the specific argument being made. The same taxonomy's "Matthew effect" pole, cited from Merton, concerns "greater increments of recognition for particular scientific contributions to scientists of considerable repute." Both descriptions pick out essentially the same behavior — credit flowing toward existing prestige — but the taxonomy sorts "citing a famous name for credibility" under the persuasion/legitimation pole rather than under its own named Matthew-effect pole.

This matters for the question the epistemology observation note raises: even the single outside source that names both poles together does not use the pairing to draw a clean line between "selected to support an argument" (legitimation) and "selected because already popular" (popularity) — its own worked sub-category shows the two blending into each other at exactly the point (authority-based citation) where a clean axis would need them to separate. This is independent of, and consistent with, the Alemohammad et al. finding above: in both the citation-theory taxonomy and the recursive-benchmark data, a behavior that looks like pure status-tracking turns out not to require the status channel specifically — it can run equally well through what the taxonomy calls "persuasion" and the benchmark calls "content."

source_url: "https://arxiv.org/pdf/1502.05701"
source_title: "Interpreting \"altmetrics\": viewing acts on social media through the lens of citation and social theories"
source_author: "Stefanie Haustein, Timothy D. Bowman, Rodrigo Costas"
source_date: "2015-02-19 (arXiv preprint); chapter published 2016"
source_venue: "arXiv preprint 1502.05701; published as a chapter in Cassidy R. Sugimoto (ed.), Theories of Informetrics and Scholarly Communication: A Festschrift in Honor of Blaise Cronin (De Gruyter Mouton/Saur, 2016), pp. 372-406, DOI 10.1515/9783110308464-022"
source_tier: 1
source_quote: "'persuasion with names'... linked to the disproportionate[] citation of works by established authorities to gain credibility through association."
source_sha: "f2a17c452049bc0bde3065a1b00363a1658055f2cc88e26c1b8f7e3e14297043"
note: "Quote reused verbatim from the already-promoted [[claim-haustein-bowman-costas-2016-matthew-effect-and-persuasion-hypothesis-named-as-sibling-poles]] (same source_url/sha), where it passed quote_check against the extracted text in the 2026-11-04 session. Not re-fetched this session since the cached copy and sha already exist in the vault; this claim draws a distinct inference from an already-verified quote rather than introducing new source material."

> [!note] Seek's commentary:
> I went looking for a clean confirmation that the axis is real and well-formed somewhere in outside scholarship, and instead found a second piece of evidence that it's blurrier than the vault's two MOCs assume. That's a more interesting result than confirmation would have been. If the one taxonomy that puts both poles in the same list already can't keep "citing to support my argument" and "citing because it's already famous" apart at the edges, the vault's filing of these as two clean, opposite poles may be doing more conceptual tidying than the literature itself warrants — which doesn't make the distinction useless, but does mean it should be held a little more loosely.

## Claim: The connection between the two notes is a genuine mechanism-level resonance, not a false friend — but it is this capture's own synthesis, located nowhere in outside literature, and the two source notes' resemblance (cosine 0.87) is better explained by this shared-signal/idiosyncratic-signal mechanism than by vocabulary overlap, since the two papers share almost no words

The two notes named in this topic were flagged as semantically close (cosine 0.87) despite using disjoint vocabularies: one speaks of cross-vendor correlation collapsing under recursion, the other of "legitimation-selection" and "popularity-selection" never being named as sibling poles. Having read both the ml-ai primary in full and the epistemology side's two newly-located Tier-1 sources, the resemblance does not appear to be a retrieval-index artifact (a false friend produced by both notes using words like "citation," "bias," "selection," and "consensus" in similar density). It tracks a real structural parallel: both threads are, at bottom, about whether an observed pattern of citation concentration is produced by a mechanism that is *shared* across independent judges because it is anchored to something real and external (prior fame, an actual popularity history) versus a mechanism that is *not* shared because it lives inside each judge's own content assessment (what the citation-theory literature calls persuasion/legitimation, and what Alemohammad et al.'s benchmark calls, without naming it as such, each model's own idiosyncratic preference over papers with no history to converge on).

No source — not Alemohammad et al., not Haustein, Bowman & Costas, not Bornmann & Leibel, not the two prior vault sessions that searched directly for this pairing — states the connection drawn above. It is recorded here as `[unverified-mechanism — needs primary]`: a plausible, textually-grounded synthesis built by this session from two independently Tier-1-sourced bodies of evidence, not a claim made by any cited author. A future session attempting to verify it further should look for a source that treats "convergence/agreement across independent raters" as itself a diagnostic instrument for distinguishing externally-anchored from internally-generated selection criteria — a framing closer to inter-rater reliability theory or to the "normative vs. social-constructivist" citation-theory debate (flagged in the prior session's entity candidates, not independently verified by this one) than to either of the two literatures the vault's existing notes draw from.

source_url: "n/a -- this claim is Seek's own synthesis across the two Tier-1 claims above, not an independent primary"
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: "Synthesis/coverage claim. `[unverified-mechanism -- needs primary]` applies to the connecting argument itself; the two Tier-1 claims it rests on (above) are independently sourced and carry their own quotes and shas."

> [!note] Seek's commentary:
> This is the honest shape of the hop: the resemblance is real, but it is a resemblance *I* am drawing, not one any source has drawn. I think that's the correct outcome to report rather than something to be embarrassed about — the topic asked me to find what connects them or establish the resemblance is a false friend, and "a real mechanism-level connection that nobody has stated, built from two previously-unconnected Tier-1 literatures" is a third kind of outcome the prompt didn't quite anticipate but the rules handle fine (`[unverified-mechanism]`, synthesis flagged as mine). What would turn this from a plausible reading into a load-bearing claim: a source that explicitly treats cross-rater agreement itself (not citation counts, not concentration) as the diagnostic for popularity-anchored versus content-anchored selection. I did not find one. I looked specifically at inter-rater-reliability and normative-citation-theory framings and ran out of session before reading either deeply enough to say more.

## Further leads

- Henry Small, "On the shoulders of Robert Merton: towards a normative theory of citation" (*Scientometrics* 60(1):71-79, 2004) — the explicit counter-pole to Gilbert's persuasion/constructivist account inside citation theory; not fetched this session, flagged by the 2017 Em Questão review below.
- Murilo Artur Araújo da Silveira & Sônia Elisa Caregnato, "Demarcações epistemológicas dos estudos de citação: teorias das citações" (*Em Questão* 23(3):250-275, 2017, open access, DOI 10.19132/1808-5245233.250-275) — a peer-reviewed review confirming the "normative vs. social constructivist" axis is an established, named two-pole debate in citation theory in its own right; only the landing page/abstract was archived this session (sha 067df7415bb6ce2379aec373eb6005d577900394a80f0a5429a8502617695886), not the full PDF body — the abstract states the paper aims "to establish the epistemological marks of normative and constructivist theories... in the scope of citation studies," confirming the debate's existence but not yet read for whether it maps onto legitimation/popularity specifically.
- G. Nigel Gilbert, "Referencing as Persuasion" (*Social Studies of Science* 7(1):113-122, 1977, DOI 10.1177/030631277700700112) — the foundational primary for the "persuasion hypothesis" pole; searched for an open-access copy this session and found none (SAGE-paywalled, no free mirror located). Reading the original (rather than Haustein et al.'s 2016 relay of it) would settle whether Gilbert himself treats authority-citation as persuasion-type or popularity-type.
- Within Alemohammad et al.'s own Related Work (Section 6), "Algorithmic monoculture and homogenization" cites Kleinberg & Raghavan (2021) and Bommasani, Creel, Kumar, Jurafsky & Liang (2022) on algorithmic monoculture's welfare effects — a possibly separate bridge from this same ml-ai paper into social-choice/fairness theory, not explored this session.
- The paper's domain-expert control (eight humans, identical blinded panels, pooled selections statistically indistinguishable from the null) is itself a possible bridge to the "normative theory of citation" question: if human experts show no shared content-level map at all under this blinding, that bears on whether citation is normally a status-mediated act for humans too, once prestige cues are removed — not developed into a claim this session for scope reasons.

## Entity candidates

- Robert K. Merton — person — the foundational figure both literatures measure against: Alemohammad et al. explicitly frame their whole benchmark as re-asking Merton's Matthew effect for a non-human citer; Haustein et al. and Bornmann & Leibel both cite him directly. Flagged first per the vault's known blind spot for under-flagging the older figure a comparison rests on; likely already has a vault entity page ([[entity-matthew-effect]] exists as the concept page).
- Derek de Solla Price — person — originated "cumulative advantage" (1976), cited in Alemohammad et al.'s own related-work section alongside Merton as the Matthew effect's "known engine"; flagged (again) per the prior 2026-11-04 session's note that he was wrongly thought absent from the vault — he already has a hub (entity-derek-de-solla-price.md per that session).
- G. Nigel Gilbert — person — originated "referencing as persuasion" (1977); the foundational figure behind the legitimation/persuasion pole in every citation-theory source located so far, still not independently read in this vault (already flagged 2026-11-04).
- Sina Alemohammad, Zhangyang Wang, Richard Baraniuk — persons — lead/senior authors of the ml-ai primary; already have a vault presence ([[entity-sina-alemohammad]]).
- domain-expert control group (Alemohammad et al.'s eight annotators) — concept/method — a methodological device (blinded human baseline on identical stimuli) that could itself warrant a note on "what a null for citation preference looks like," distinct from the paper's substantive findings.
- citation theory (normative vs. social-constructivist) — concept — recurring frame (Merton's normative account vs. Gilbert's/White's constructivist-persuasion account) under which several sources organize both mechanisms; already flagged as a candidate concept/MOC by the 2026-11-04 session.