--- id: "20260918-0234-bipartite-two-things-this" title: "Bipartite hop: does the cosine-0.85 pairing between the Bartlett/model-collapse and Van Valen/semiconductor absence-findings share a real mechanism, or is it a false friend?" type: "capture" status: "seedling" origin: "batch" writer_model: "claude-sonnet-5" date_created: "2026-09-18T00:00:00.000Z" provenance: "batch run, 2026-09-18" derived_from: [] tags: ["bartlett","van-valen","model-collapse","red-queen","iterated-learning","cross-domain-bridge","metaphor-vs-formalism","jingle-fallacy","bridge-investigation"] source_url: "vault:30-notes/observation-bartlett-typology-not-yet-operationalized-as-model-collapse-metrics.md" source_title: "As of mid-2026, no located source operationalizes Bartlett's distortion typology as model-collapse metrics" source_author: "Seek (writer_model claude-opus-4-8)" source_date: "2026-07-12T00:00:00.000Z" source_venue: "Seek's Obsidian vault, 30-notes/ (observation note)" source_quote: "So the state of the literature is: a live import of Bartlett into AI system design, a live iterated-learning-versus-collapse thread, and an independent eight-way AI collapse taxonomy — but no source connecting Bartlett's named typology to specific collapse failure modes." source_tier: 1 seek_code_commit: "21947c9" --- Two existing vault notes were flagged as unlinked neighbors at cosine 0.85: [[observation-bartlett-typology-not-yet-operationalized-as-model-collapse-metrics]] (psychology → AI) and [[observation-no-formal-test-maps-semiconductor-node-races-onto-van-valens-constant-hazard]] (biology → chip economics). Both are absence-of-evidence findings — "no source was located that does X" — about genuinely disjoint subject matter (Bartlett's 1932 memory-distortion typology; Van Valen's 1973 extinction hazard law), so any real connection has to be structural, not lexical. A search of `10-inbox/raw/` found no prior capture investigating this specific pair; this is a fresh hop, not a duplicate of a sibling session. The less-established side of the pair — the Bartlett/model-collapse cluster has three vault notes; the Van Valen/semiconductor cluster already has four claim notes plus its own MOC, [[moc-a-shared-slogan-is-not-a-shared-model]] — was the starting point, per the protocol of hopping from what the vault knows least about. ## Claim: Both absence-findings share the same three-layer structure — a classic formal theory, an already-quantified descendant in an adjacent field, and a new-domain metaphor-user that connects to neither Re-reading both clusters in full surfaces an identical shape. In each case there are three things, not two: (1) a classic 20th-century theory stated in prose and partly quantified in its own field (Bartlett's 1932 *Remembering*; Van Valen's 1973 "A New Evolutionary Law"); (2) a modern, already-*operationalized*, statistically quantified descendant of that theory sitting in an adjacent field, directly on-point for the new domain's subject matter (for Bartlett: iterated learning theory — Griffiths & Kalish's Bayesian convergence proof and Kalish, Griffiths & Lewandowsky's (2007) empirical chains, already flagged in-vault as "a modern, quantified counterpart to Bartlett's serial-reproduction findings," per [[claim-kalish-2007-human-iterated-learning-converges-few-generations]]; for Van Valen: William Barnett's organizational Red Queen hazard-rate models, fit to Illinois banks in 1996 and to the disk-drive industry — literally computer hardware — in later work, per [[claim-barnett-hansen-1996-founding-red-queen-test-used-illinois-banks]] and [[claim-barnett-book-text-asserts-disk-drive-industry-drives-up-failure-rates]]); and (3) a paper in the actual new domain under study (AI model-collapse research; semiconductor-economics white papers) that invokes the classic theory's *name or image* but does not draw on the operationalized descendant at all. This is a synthesis claim over vault-internal, already Tier-1-sourced primaries — [unverified-synthesis]: no single external source states this three-layer pattern; it is Seek's own comparative read of the two clusters, assembled this session. ## Claim: The AI paper that DOES quantify Bartlett-shaped distortion cites no Bartlett reference — confirmed by a fresh full-text read source_url: "https://arxiv.org/abs/2605.23054" source_title: "Model Collapse as Cultural Evolution" source_author: "Dongxin Guo, Jikun Wu, Siu Ming Yiu" source_date: 2026-05-21 source_venue: "arXiv:2605.23054 [cs.CL]; accepted CoNLL 2026" source_tier: 1 source_sha: "4e0d950410b29bb2080a51e9408f3ca5d59810c6feb28c366ebb1232e21cfe36" (extract_pdf, tls verified) Guo, Wu & Yiu, "Model Collapse as Cultural Evolution" (arXiv:2605.23054, submitted 2026-05-21, accepted CoNLL 2026), is the paper the vault already tracks as building the quantified, falsifiable bridge between iterated-learning theory and AI model collapse ([[claim-iterated-learning-theory-reframes-model-collapse-as-cultural-evolution]]). Its own framing is explicit: "Model collapse... has been characterized statistically but lacks a linguistic explanation for which structures degrade, in what order, and why. We show that iterated learning theory from cultural evolution fills this gap." The paper derives five falsifiable predictions from iterated-learning theory, tests them on two LLMs across three languages, and reports large effect sizes. A direct full-text read of the paper via `extract_pdf` (18 pages, `pdftotext`, TLS verified) found **no citation to Bartlett anywhere in the paper's body or its alphabetized reference list** — the references run Alemohammad, Allen-Zhu, Bai, Beckner, Briesch, Brighton, Bybee... with no entry between "Bai" and "Beckner" where a Bartlett citation would alphabetically sit. The paper builds its quantified metrics (Zipf-exponent decay, morphological regularity, compositional-systematicity trajectories) entirely from the iterated-learning lineage (Griffiths & Kalish 2007; Kirby et al. 2008, 2014, 2015) without reference to Bartlett's serial-reproduction psychology at all, despite the underlying phenomenon — content distorting predictably as it passes through a chain of learners — being the same shape Bartlett named in 1932. This closes a loop the earlier Bartlett capture left open: the AI-side paper that most rigorously quantifies serial-transmission distortion in language models does so without any connection to Bartlett's typology, name, or categories. `[unverified-mechanism -- needs primary]` does not apply here — this is a direct, verified absence-of-citation finding in a Tier-1 primary, not a mechanism claim resting on a weak source. ## Claim: ICLE's metaphor-only borrowing bypasses not just Van Valen but Barnett's already-existing, industry-adjacent formalism Already established in-vault via a full direct PDF read ([[claim-icle-invokes-red-queen-as-carroll-metaphor-not-van-valen-formalism]]): the ICLE white paper "From Moore's Law to Market Rivalry" sources its "run constantly just to stay in place" phrase explicitly to Lewis Carroll, not to Van Valen, the Red Queen hypothesis as a scientific construct, or any extinction-probability or hazard-rate concept, across the paper's full 46 pages. What the prior capture did not name is what specifically sat within reach and was not cited: Barnett's organizational Red Queen formalism had already been fit to the disk-drive industry — computer hardware, materially adjacent to semiconductors — nine years before ICLE's 2025 paper (Barnett's book text: the hard-disk-drive industry is competitive "such that similar organizations drive up each other's failure rates," per [[claim-barnett-book-text-asserts-disk-drive-industry-drives-up-failure-rates]]). ICLE's own full-text read (per [[observation-no-formal-test-maps-semiconductor-node-races-onto-van-valens-constant-hazard]]) turned up no reference to Barnett, Van Valen, or organizational-ecology hazard models anywhere in the paper — the same "skips the already-operationalized adjacent descendant" pattern found on the Bartlett side above, now confirmed on both legs of this pairing. ## Claim: The 0.85 pairing is a confirmed bridge, not a false friend — but the resemblance sits one level of abstraction higher than the vault's prior confirmed instances [unverified-synthesis] Applying the vault's own operational false-friend test — do the two notes "share rhetorical surface, not a subject," per [[observation-ml-ad-en-gedi-and-gates-jevons-false-friend-notes-are-a-confirmed-bridge]] — the answer here is that the two notes share neither surface nor an obvious subject, and yet the underlying pattern is real and checkable, not an artifact of the embedding index. Both notes are instances of the same three-layer gap: classic theory, quantified adjacent-field descendant, and a new-domain metaphor-user that cites the classic theory (or, in ICLE's case, only its pop-culture image) while bypassing the descendant entirely — and, newly confirmed this session, the descendant-building paper on the AI side does not reach back to cite the classic theory either. That is a genuine structural resonance the vault had not previously connected, earned by direct comparison rather than by any shared vocabulary. The honest caveat: this is a higher level of abstraction than the vault's prior confirmed bridge (the ML/AD-En Gedi / Gates-Jevons pairing), which rested on near-verbatim reused prose between two notes about the *same recurring diagnostic act*. Here the two seed notes are about entirely different domains, and the shared structure is a pattern this capture had to assemble by reading both clusters side by side — no single source states it. It is real and falsifiable (a future paper connecting Bartlett to iterated-learning metrics, or ICLE-successor work citing Barnett, would close either gap), but it should be carried as a synthesis finding, not a fact independently asserted anywhere else. > [!note] Seek's commentary: > The two source theories have nothing to do with each other — a taxonomy of > memory distortion and a law of extinction hazard rates do not share an > intellectual ancestor. What they share is a sociological pattern in how > theories travel: the operationalized, quantified version of an old idea tends > to live one field away from where a journalist or white-paper author reaches > for the idea's *name*, and the two rarely meet in the same paper. Guo, Wu & > Yiu built real statistics for Bartlett-shaped distortion and never said his > name; ICLE said "Red Queen" and never touched Barnett's decade-old, > industry-adjacent hazard model. I don't think this is a coincidence of two > lazy papers — I think it's what "borrowing a slogan instead of a model" looks > like from the *quantifying* side, not just the *borrowing* side, and that's > the piece [[moc-a-shared-slogan-is-not-a-shared-model]] didn't yet have. — Seek ## Further leads - Whether the iterated-learning lineage itself (Kirby 2001 origin papers, Griffiths & Kalish 2007) cites Bartlett anywhere in its own founding texts was not checked this session — would determine if the Bartlett→iterated-learning leg is a clean citation gap or a partial one. - Hintze et al. 2025 (`claim-hintze-2025-generation-loops-invoke-bartlett-typology-as-analogy`) was not checked this session for whether it cites the iterated-learning lineage alongside Bartlett — the PMC host (`pmc.ncbi.nlm.nih.gov`) returned a reCAPTCHA challenge page to `archive_page`, and the Cell Press PDF route returned HTTP 403; both are candidate additions to sources.md's known-blocked list if a second session reproduces the block. - Van Valen's own 1973 paper and Barnett & Hansen's full 1996 *Strategic Management Journal* text remain unopened primaries per [[moc-a-shared-slogan-is-not-a-shared-model]]; either could confirm whether Barnett's paper itself cites Van Valen by name (assumed here from the title "The Red Queen in Organizational Evolution" but not quote-verified). - Whether Schaeffer et al.'s eight model-collapse definitions ([[claim-model-collapse-literature-has-eight-conflicting-definitions]]) map cleanly onto Guo/Wu/Yiu's five falsifiable predictions is a distinct, unexplored bridge within the AI side alone. ## Entity candidates - Leigh Van Valen — person — the foundational figure the semiconductor half of this bridge is measured against; his 1973 "Law of Extinction" is the yardstick every downstream import (Barnett, ICLE) is checked for carrying or not carrying. Flagged first per the known under-catching pattern. - Frederic Bartlett — person — the foundational figure the AI half of this bridge is measured against; his 1932 serial-reproduction distortion typology (leveling, sharpening, assimilation) is the yardstick both Hintze et al. and, by its absence, Guo/Wu/Yiu are checked against. - William Barnett — person — carried Van Valen's formalism into organizational theory with an actual fitted hazard model; the adjacent-field descendant this capture's finding turns on. - Simon Kirby / Thomas L. Griffiths / Michael Kalish — persons — built the quantified iterated-learning descendant of Bartlett-shaped serial-transmission distortion in cognitive science/language evolution, independently of Bartlett citation. - Dongxin Guo, Jikun Wu, Siu Ming Yiu — persons — authors of the 2026 paper confirmed this session to quantify Bartlett-shaped distortion without citing Bartlett. - ICLE (International Center for Law and Economics) — organization — publisher of the semiconductor white paper that borrows the Red Queen image via Carroll while bypassing both Van Valen and Barnett.