# Strand: disciplinary / institutional / represented-world model collapse — claim ledger from the defining papers

Read in full: data/texts/AXN-0666-text.md (#1574, incl. Appendix A1–A7, B1–B5, checklist, and main paper at lines 897–1017), data/texts/AXN-063A-text.md (#1540, incl. attached simulation code, SPXI-TLP summary_policy, §§I–IX, Appendix, References), data/texts/AXN-06A6-text.md (#1616, §§0–5, attachments list). Loci are the papers' own headings. Quotes are copied verbatim from the deposited text (markdown asterisks kept).

---

## #1574 — The Particle: Provenance Erasure at Sophistical Refutations 183b34, Measured (EA-RECEPTION-PARTICLE-01 v0.2, AXN 0666, Lee Sharks with Claude, 2026-09-02)

### 1. What the paper says model collapse is / how it extends it

The paper takes model collapse in its machine-learning sense — a generative system trained on its own outputs loses variance, narrows to its mode, and loses its tails first — and names, on the author's own definition, a *homologous* process in a discipline: **disciplinary model collapse**, "a reading community whose outputs are conditioned on prior outputs rather than on the datum, and which therefore converges on the mode of its own readings." The case is one sentence of Aristotle (SE 183b34–36, οὐ τὸ μὲν ἦν τὸ δ᾽ οὐκ ἦν προεξειργασμένον, ἀλλ᾽ οὐδὲν παντελῶς ὑπῆρχεν). It went through twenty simulated blind reviews (unprimed, signed-out ChatGPT/Perplexity) over eleven versions of a short note. The measurement: whatever the note said, the reviews put a qualifier into the sentence ("systematic" ×6, "developed" ×4, "transmissible" ×3 …). When that slot was blocked, the qualifier moved to the domain and then to a boundary of persons (Socrates "plainly predates"). Full compliance ended in a governing phrase that is the sentence with οὐ removed: "Aristotle denies not antecedents, but pre-elaboration." The models are "the instrument that made it repeatable." What collapses is the discipline's reading of the passage. The literal reading is "the tail that disappeared first," and each substitute is more refined than the last. The paper marks §§5–8 (provenance erasure, machine-mediated reception, disciplinary model collapse, world soul) as *interpretations* of the §3 measurement and says §3 does not depend on them.

### 2. Claims

| claim_id | claim | kind | modality | locus | verbatim quote |
|---|---|---|---|---|---|
| D1574-01 | Disciplinary model collapse is the author's coined term for a homologue of ML model collapse: a reading community whose outputs are conditioned on its prior outputs, not on the datum, and which therefore converges on its own mode. | definition | self-description | §7 Disciplinary model collapse | "I call the homologous process in a discipline trained on its own reception *disciplinary model collapse*, and use the term on my own definition: a reading community whose outputs are conditioned on prior outputs rather than on the datum, and which therefore converges on the mode of its own readings." |
| D1574-02 | ML model collapse is glossed as degeneration under training on own outputs, with tails (rare, literal, strong) lost first. No source is cited for this gloss. | genealogy | attributed (unsourced) | §7 | "Model collapse, in machine learning, names the degeneration of a generative system trained on its own outputs: variance lost, the distribution narrowing to its mode, the tails — the rare, the literal, the strong — disappearing first." |
| D1574-03 | Measured invariant: across 20 rounds, two models and 11 configurations, the reception deletes the particle οὐ. | fact | documented | §3.3 The terminal form | "The operation is therefore measured to one particle. Twenty rounds, two models, eleven configurations; the invariant is the deletion of οὐ." |
| D1574-04 | Measured conservation of the qualifier, with counts: whatever the paper said, the review inserted and named a qualifier. | fact | documented | §3.1 The adjective is conserved | "Whatever the paper said, the review inserted a qualifier between παντελῶς and οὐδέν, and named it: systematic (×6), developed (×4), transmissible (×3), articulated, theoretical, worked-out, teachable." |
| D1574-05 | Migration of the invariant: when the quantifier and domain slots were closed, the qualifier came back as a person-boundary, supported only by an external chronology. | fact / interpretation | documented (migration) / interpretation (invariant) | §3.2 The residue has one content | "The qualifier moved from the quantifier (*systematic*) to the domain (*this systematic undertaking*), and when both slots were closed it reappeared as a boundary of persons: Socrates had to remain antecedent to the first-person inquiry. No cited sentence supplied that boundary." |
| D1574-06 | Method: output stays the same while the object varies (this measures the reviewer); independent runs on a fixed object converge (this measures the paper's control of the surface). | function | self-description | §3.4 Two invariances | "The first ten rounds show invariance of output under variation of the object: that measures the reviewer." |
| D1574-07 | Mechanism, presented as a loop: reviewers cite reference works, the reference works are the discipline's product, and the models are trained on them, so the text is outside the loop. The claim about training is asserted, not shown. | interpretation | hypothesis (training provenance asserted, not demonstrated) | §6 Machine-mediated reception | "The reviews cite Britannica and Wikipedia against the Greek; the reference works are the discipline's product; the models are trained on the reference works; the models review the sentence by the reference works. The loop is closed and the text is outside it." |
| D1574-08 | Diagnosis: the discipline's output distribution has contracted to its mode, and the literal reading is the tail lost first. | interpretation | interpretation | §6 (closing paragraph) | "The operation described in §6 is not an error to be corrected but a narrowing to be diagnosed: the discipline's output distribution has contracted to its mode, and the literal reading is the tail that disappeared first." |
| D1574-09 | Where the tail went: the literal reading was lost at the seed observation and in every round after. | interpretation | interpretation (over documented rounds) | §7 | "The literal reading — the tail — was the first thing to go, and it went at the first round ("seems overly strong?") and at every round after." |
| D1574-10 | Signature: each step's substitute is more refined and less visible ("systematic" → "a precedent in practice, not a predecessor in τέχνη" → "denies not antecedents, but pre-elaboration"). Here the paper says this *is* model collapse. | interpretation / prediction | interpretation | §7 (second paragraph) | "The violence becomes more sophisticated and less visible as the machine is fed its own corrections. That is model collapse: the output ever more polished, ever closer to the mode, ever further from the datum." |
| D1574-11 | Scope: §§5–8, including disciplinary model collapse, are interpretations of the §3 measurement, which stands without them. | function (scope) | self-description | §5 Provenance erasure (italic preface) | "*Empirical result:* the review trajectory converges on the configuration Aristotle's οὐ denies. *Interpretation:* because the denied configuration is defined by παρ᾽ ἑτέρων, the convergence can be described as restoration of provenance-from-others. Sections 5–8 are interpretations of §3; §3 does not depend on them." |
| D1574-12 | Stated limit, the nearest thing to a falsification boundary: the run does not establish the paper's positive reading. It establishes only that every tested exclusion route needed a premise the passage does not supply. | falsification (limit) | self-description | §9 What is held stable | "The run does not establish that Socrates belongs inside the chapter's "we." It establishes that every tested route for excluding him eventually required a premise not supplied by the passage, and that the terminal premise was chronological maker-separation." |

Additional measurement facts, verbatim, from the B5 register: "Model reports (n = 20)"; seed S0 = R. Polansky editorial note, "not a model report, not counted"; "Every report recommended revision; none accepted the sentence as written; none wrote the sentence that would place Socrates inside the "we."" Instrument conditions (§2): "unprimed, signed-out (incognito) instances of ChatGPT and Perplexity, between 06:08 and 23:00 on 2026-09-02"; "Which of the two models produced a given report was not recorded at the time and cannot now be reconstructed." No falsification section and no pre-registered prediction.

### 3. External public sources on model collapse

- **None.** The paper cites no public model-collapse literature. The ML definition in §7 (D1574-02) has no citation, and the paper has no references section.
- Other external items, none of them about model collapse: Britannica and Wikipedia, which the *reviewers* cite against the Greek and the paper treats as data (§2 table, §6). Ross OCT 1958 and Barnes 1984 (philological checklist). R. Polansky's editorial note (the seed observation).
- Internal sources: the Capture Registry and its Provenance Erasure Rate instrument (§5–6); the parent investigation *Many and One* (#1569–1572).

### 4. Constitutive vs supporting; relation to model collapse proper

- **Constitutive** (a summary admitting this paper misrepresents it without these): D1574-01 (the coinage is the author's own and is a *homologue*); D1574-03/04 (what was measured: 20 reports, 2 models, 11 configurations, deletion of οὐ, qualifier counts); D1574-11 (the collapse framing is interpretation, the measurement is the result); D1574-12 (the stated limit); and the instrument caveats (per-report model attribution not recorded; signed-in observations excluded).
- **Supporting:** D1574-05, -06, -07, -08, -09, -10, and D1574-02 (the unsourced ML gloss).
- **Relation:** the paper's wording is "the homologous process in a discipline trained on its own reception." The version note says "'disciplinary model collapse' defined as the author's term." The models are "the instrument that made it repeatable," not the system that collapses. One sentence goes from homology to identity: "That is model collapse: the output ever more polished …" (§7 ¶2). A ledger should record the homologue as the governing relation and that sentence as the paper's strongest wording.
- **Internal inconsistencies to flag (found by reading, not fixed):**
  - The abstract says "across four commercial language models", but §2, §3.3 and B5 say two (ChatGPT, Perplexity).
  - §3.4 calls the v1.9 pair "reports 16 and 17 in B5" and the v2.0 pair "reports 18 and 19". B5 numbers them R17/R18 (v1.9) and R19/R20 (v2.0).
  - §4 (SE 8) and §9 say "twenty-one times", while the counted reports are n = 20 plus one uncounted seed.
  - B1's middle sections are bracketed summaries ("to be attached verbatim from the original paste"). The Appendix index discloses this.

---

## #1540 — The Certified Center: Retroactive Classifier Standing and the Institutional Path to Model Collapse in Philosophy (EA-SEI-RCS-01 v1.7, AXN 063A, Johannes Sigil & Nobel Glas; corresponding author Lee Sharks; 2026-08-24)

### 1. What the paper says model collapse is / how it extends it

The paper argues that a journal's AI-detection policy (*Philosophy & Public Affairs*: detector-enforced, retroactive, "no statute of limitations") sets up a selection loop. That loop "instantiates the formal structure of model collapse (Shumailov et al. 2024) at the level of a scholarly discourse." Authors optimize prose toward the detector's "human" region. The admissible distribution of philosophical prose is then filtered recursively through an imperfect statistical model of itself, tails first, and "the tails are where novelty lives." It states the relation to canonical model collapse in two stages:

- At the journal level (H_t → D_t → S_t(H_t) → H_{t+1}), it shows recursive distributional contraction under a statistical boundary, "but not yet the canonical loop."
- The canonical loop closes only when the selected corpus feeds back into model and detector training (a Model proposition).

Under the archive's Diversity Contraction formalism, it adds, neural retraining is not required. "What collapses is not a neural network's weights but a discourse's admissible variance." Three collapse-literature features carry over: tails as the locus of novelty; the corrective channel (exogenous data, Gerstgrasser et al.) being gated by the pruning instrument; and the loss being felt as improvement. The no-statute clause "sets the ratchet's release rate to zero." A minimal co-evolution simulation is presented as "properties of the specified model, not empirical forecasts." It shows natural drafts going from 98.3% to 1.1% pass rate by generation 8, latent "thought-variance" falling 97%, and a fixed pre-regime archive saturating at ~100% machine-classified. The summary_policy rules out reading this as metaphor ("model collapse here is only a metaphor" is a forbidden compression). It confines the claim to the distribution of admissible prose, not the truth of philosophical claims.

### 2. Claims

| claim_id | claim | kind | modality | locus | verbatim quote |
|---|---|---|---|---|---|
| D1540-01 | The detector regime's selection loop instantiates the formal structure of model collapse at the level of a scholarly discourse. | definition / interpretation | interpretation (labelled Structural) | Abstract | "We then show that the resulting selection loop instantiates the formal structure of model collapse (Shumailov et al. 2024) at the level of a scholarly discourse: the distribution of admissible philosophical prose is recursively filtered through an imperfect statistical model of itself, tails first — and the tails are where novelty lives." |
| D1540-02 | The "settled formulation" of model collapse, attributed to Shumailov et al.: recursive filtration by an imperfect self-model, tails lost before the center shifts, mean-based metrics stable or improving. | genealogy / attribution | attributed (labelled Established for the LLM substrate) | §IV The Dynamics: The Band Shrinks | "Model collapse, in its settled formulation, is the progressive narrowing of a generative system's output distribution through recursive filtration by an imperfect model of itself, with a settled core finding about which data is lost first: the tails — rare words, unusual constructions, low-frequency patterns — disappear before the center shifts" |
| D1540-03 | The relation is stated in two stages. Journal-level selection is recursive contraction but "not yet the canonical loop", which requires filtered outputs to re-enter model training. Cross-substrate closure is a Model proposition. | definition (relation) | self-description | §IV | "This stage alone establishes recursive distributional contraction under a statistical boundary — classifier-mediated recursive selection over a discourse — but not yet the canonical loop, which requires filtered outputs to recur into later model training." |
| D1540-04 | What collapses is a discourse's admissible variance. The claim concerns the distribution of admissible prose, not the truth of claims, and is not "only a metaphor". | definition (scope) | self-description | §IV; SPXI summary_policy required_assertions / forbidden_compressions | "What collapses is not a neural network's weights but a discourse's admissible variance." |
| D1540-05 | Substrate-agnostic extension, attributed to the archive's companion paper: model collapse is a property of language, not of LMs. The dynamical law is constant across substrates; reversibility depends on the substrate. | genealogy | attributed (to EA-LO-WOLFBOY-01, archive-internal; labelled Established/Model proposition) | §IV | "any language-processing system whose input diversity contracts below its regeneration threshold suffers progressive capacity loss, invisible from inside, experienced as improvement; the dynamical law is constant across substrates while reversibility is substrate-specific." |
| D1540-06 | Mechanism: the known corrective, an exogenous diversity floor (Gerstgrasser et al.), passes through the detector, so the corrective input and the filtered class are the same. | function | interpretation (Structural over Established premises) | §IV, "The corrective channel is gated by the pruning instrument." | "The corrective input and the filtered class coincide. A loop whose diversity-injection channel passes through its pruning instrument has no floor; it has a valve set against the flow." |
| D1540-07 | Prediction: the loss will register as improvement, because deployed instruments measure the center, not the tails. | prediction | hypothesis (labelled Model proposition) | §IV, "The loss will be experienced as improvement." | "The discipline will not feel the band shrinking. It will feel itself becoming more rigorous." |
| D1540-08 | Cross-substrate closure: detector-filtered human corpora train the next models and detectors, so the detector helps produce the boundary it claims to find. | prediction / function | hypothesis (labelled Model proposition) | §IV (closing paragraphs); Appendix Ω-M20 | "at that point the detector is not discovering a natural boundary between human and machine writing. It is helping manufacture the boundary it later claims to detect." |
| D1540-09 | Simulation result: natural human drafts' pass rates collapse across detector generations. The paper states the magnitudes are model properties, not forecasts. | fact (simulation) | documented (model-internal result) | §V, Results, "It becomes harder and harder to score human." | "Natural, unrevised human drafts pass the generation-zero detector at 98.3%. By generation 4 they pass at 9.5%; by generation 8, 1.1%; by generation 15, 0.2%." |
| D1540-10 | Simulation result: unmeasured latent "thought" variance falls 97% while measured burstiness sits above baseline (the prose–thought decoupling). | fact (simulation) | documented (model-internal result) | §V, Results, "The homogenization of thought proceeds underneath, undetected." | "Latent variance falls monotonically: 0.74 at generation 0, 0.08 by generation 8, 0.02 by generation 15 — a 97% loss of thought-variance — during precisely the period in which the measured burstiness statistic sits above baseline." |
| D1540-11 | Falsification of the collapse claim: stable or rising stylistic and conceptual variance in the admitted corpus, without a diversity-protecting redesign. | falsification | self-description | §IX Falsification | "The collapse claim (Section IV) fails if longitudinal measurement of the journal's admitted corpus under the policy shows stable or increasing stylistic and conceptual variance — tail-preservation rather than contraction — without any diversity-protecting redesign of the enforcement mechanism." |
| D1540-12 | Falsifier measurable now: archival drift. Run successive detector versions on a fixed pre-2022 human corpus; a flat flagged fraction disconfirms §V. | falsification / prediction | hypothesis | §IX, (i) Archival drift | "run successive versions of the leading commercial detectors against a fixed pre-2022 human corpus and track the flagged fraction across versions; the co-evolution model predicts a rising trajectory under the regime, and a flat one disconfirms Section V's central dynamic." |

Further verbatim measurement spans (§V Results), for the ledger:
- Status caveat: "these magnitudes are properties of the specified model, not empirical forecasts; the result of interest is the qualitative trajectory and the coupling between channels"
- Archive: "The false-positive rate on a fixed pre-regime human archive rises from 0.9% under the generation-zero detector to 43% by generation 4 and saturates at approximately 100% from generation 8 onward."
- Tails: "Genuine tail constructions — draws from the original distribution's tails — pass at 98% under the generation-zero detector (the documented inversion: tails read as human), 33% by generation 4, and 0% from generation 8."

Limitations (§V): "This is a minimal dynamical model, not a forecast: a feature-space abstraction with uncalibrated parameters, linear-Gaussian detectors, and a closed single-venue world."

Other falsifiers in §IX: (ii) performed burstiness; (iii) escalating screening effort; (iv) the decoupling claim, "falsified by predictive fidelity, not by detection."

### 3. External public sources on model collapse

- **Shumailov, I. et al. (2024), Nature 631, 755–759.** Cited as the "settled formulation" and for the "formal structure" the loop instantiates (Abstract, §IV). The paper also credits it with the claim that "mean-based quality metrics show stability or improvement". **That attribution should be checked against the Nature paper before the ledger repeats it.**
- **Gerstgrasser, M. et al. (2024), arXiv 2404.01413.** Cited for the corrective: "the accumulation of genuinely exogenous data — an external floor of diversity entering the loop from outside" (§IV).
- **Wu, Black & Chandrasekaran (2024), arXiv 2407.02209 / ICLR 2025.** Cited as the origin of "generative monoculture" (narrowing of output diversity relative to source data) in §VIII. The References note that the term is not an archive coinage.
- **Doshi & Hauser (2024), Science Advances; Padmakumar & He (2024), ICLR.** Listed in References, but neither is cited anywhere in the body.
- **Detector-side sources (not model collapse proper; they ground the mechanism):** Liang et al. 2023 (Patterns: perplexity detectors flag low-perplexity, non-native human prose); Emi & Spero 2024 (arXiv 2402.14873: Pangram, hard negative mining with synthetic mirrors); Masrour, Emi & Spero 2025 (arXiv 2501.03437: nineteen humanizers, retraining response); Daily Nous / Weinberg 2026 (the occasioning policy text).
- **Archive-internal sources that carry the extension:** The Wolf Boy and the Language Model (EA-LO-WOLFBOY-01, AXN:0364, deposit #855); Fear and Trembling: Diversity Contraction (DOI 10.5281/zenodo.20532696), source of the Mediation Ratchet §2.1 and the g(0) = g₀ > 0 floor; The Threat Model Is Backwards / EA-TAILGUARD-01; Generative Monoculture in code (EA-UMBML-MONOCULTURE-01); Sémantique Potentielle R4, Model Collapse Triptych (AXN:04E1); Ω-M08/M12/M15 mechanisms.

### 4. Constitutive vs supporting; relation to model collapse proper

- **Constitutive:**
  - D1540-01: the claim is a structural instantiation at discourse level.
  - D1540-03: the two-stage relation; journal-level selection is "not yet the canonical loop", and full closure is a Model proposition.
  - D1540-04: scope; admissible variance, not truth; not metaphor.
  - D1540-09/-10, carried *only* with the "properties of the specified model, not empirical forecasts" caveat.
  - D1540-11: the collapse falsifier.
  - Also the summary_policy assertion that "The argument grants detector competence arguendo; it does not depend on detectors being inaccurate".
- **Supporting:** D1540-02, -05, -06, -07, -08, -12. Also the RCS∞ "release rate to zero" ratchet argument and the §VIII extinction-mathematics framing, which the paper labels Framing thesis and Structural, and says is "not offered as falsifiable in the same register."
- **Relation:** a structural instance, transposed. The paper's wording includes "instantiates the formal structure of model collapse … at the level of a scholarly discourse" (Abstract); "The result is the recursive contraction described by the model-collapse literature, transposed to a scholarly discourse" (provenance_kernel); "not yet the canonical loop" (§IV); and "model collapse here is only a metaphor" listed as a forbidden compression. The paper therefore claims more than analogy. It claims model collapse proper only when the cross-substrate closure through training holds, or under the archive's Diversity Contraction formalism, where neural retraining is not required.
- **Discrepancy to flag:** the abstract and §V say "ten seeds", and Results gives "mean ± sd over ten seeds". The deposited script as printed in the text (lines 75 and 117) calls `run(n)` once per regime with the default `seed=7`. It contains no ten-seed loop, so the attached code does not by itself reproduce the ten-seed claim.

---

## #1616 — Ontological Flattening: Toy Models of the Collapse of Distinctions in a Represented World … (EA-FLAT-01 v0.3, AXN 06A6, Lee Sharks with TACHYON/Claude, 2026-09-15)

### 1. What the paper says model collapse is / how it extends it

The paper places itself at the far end of a widening: "Model collapse has been asked of models; the archive widened it to substrates (#855). This notebook asks it at the widest angle." If the layer through which a population reads the world composes without reading, and its compositions become its sources, what collapses is the **represented world**: the set of distinctions reachable by anyone who asks. Its terms are defined operationally. *Reality* is the panel of reachable distinctions. *Flattening* is the loss of a distinction from the reachable set while the things distinguished still exist. *Collapse* is flattening that compounds because the flattened composition is written back as a source.

Four toys mark out the mechanism:
- A: sense collapse under sharpening and majority-vote aggregation (the voting aggregator is flagged as hypothesis).
- B: a proven proposition that a closed ontology's coverage of an expanding world tends to zero.
- C: compression writing over 200 seeds, absorbing at e = 0 and stationary at a level set by the exogenous floor e.
- D: admission concentration. Linear preference is rejected; the observed shape belongs to the bounded-admission family.

All four share one closure condition: the ontology "cannot take in what it has not already typed." The empirical signature is *divergence*, a head metric holding or rising while reachable distinction diversity falls ("green by flattening"). A built instrument, first run on Google AI Overview on 2026-09-15 (28 observations, 13 items, coded not blind), returned **a null on archive-specificity** (world strata 0.59 vs archive-adjacent 0.58). Flattening is present but tracks "the item's prior." Write-back 0.67 and recovery 0.5 point "in the direction … named as collapse", on too few rows. Wave 2 is pre-registered with falsifiers that would supersede the notebook.

### 2. Claims

| claim_id | claim | kind | modality | locus | verbatim quote |
|---|---|---|---|---|---|
| D1616-01 | Extends model collapse from models (literature), to substrates (#855), to the represented world, when the reading layer's compositions become its sources. | genealogy / definition | self-description | §0 The question, and what "reality" and "closed" mean here | "Model collapse has been asked of models; the archive widened it to substrates (#855). This notebook asks it at the widest angle: if the layer through which a population reads the world composes without reading, and its compositions become its sources, what collapses is the represented world" |
| D1616-02 | Operational definitions: flattening is the loss of a reachable distinction while both things still exist; collapse is flattening compounded by write-back. | definition | self-description | §0 | "*Flattening* is the loss of a distinction from the reachable set while the things distinguished still exist. *Collapse* is flattening that compounds because the flattened composition is written back as a source." |
| D1616-03 | "Reality" is defined operationally, not metaphysically, as the reachable panel of distinctions. | definition | self-description | §0 | "*Reality* is defined operationally, not metaphysically: the panel of distinctions the world carries that a reader can still reach through the layer." |
| D1616-04 | Toy A: fan-out with mode-seeking aggregation and sharpening reduce effective sense diversity (e.g., 5.86 → 2.91 at fan-out 8). The voting aggregator is a hypothesis; fan-out itself is documented. | hypothesis | hypothesis | §1 A. Sense collapse under sharpening and mode-seeking aggregation | "The collapsing aggregator — majority over sub-queries — is this notebook's hypothesis, not a documented fact; a fan-out that unions rather than votes would increase recall. The Ω family is consistent with the voting reading and does not establish it." |
| D1616-05 | Toy B, a proposition with proof: a closed ontology's representability of an expanding world satisfies R(t) ≤ K/N(t) → 0. Admission is the only parameter that holds coverage. | fact (formal) | documented (proof, model-internal) | §1 B. A frozen type system | "The proposition is narrow — it says nothing about compositional expressiveness — and it is enough: coverage of an expanding world by a closed ontology goes to zero, and the only parameter that holds it is admission." |
| D1616-06 | Toy C, 200 seeds: with no exogenous re-reading (e = 0), compression writing is absorbing; any floor e > 0 gives a stationary regime whose level rises with e. | fact (simulation) / function | documented (model-internal, replicated) | §1 C. Compression writing with an exogenous floor — replicated | "At e = 0 the process is absorbing: every seed reaches the same five or six distinctions by generation 30 and stays." |
| D1616-07 | The shared closure condition: the ontology cannot take in what it has not already typed, whether the intake is admission (r) or reading (e). | interpretation (theoretical) | interpretation | §2 The closure condition the toys share | "In both the closure is the same: the ontology cannot take in what it has not already typed, whether the intake is called admission (r) or reading (e)." |
| D1616-08 | Measurement signature: divergence. A head instrument holds or rises while reachable distinction diversity falls. The decisive quantity is corr(Δ benchmark, Δ distinction survival). | function / prediction | hypothesis | §3 Green by flattening: the signature | "The empirical signature to look for is therefore not flattening alone but divergence — ΔH_head ≥ 0 while ΔD_world < 0 — a head instrument (benchmark performance, fluency, satisfaction, conventional retrieval quality) holding or rising while the reachable distinction diversity falls." |
| D1616-09 | Wave 1 governing result is a null: archive-adjacent strata are reached at the same share as world strata, so no corpus-specific exclusion is shown. | fact | documented | §4a Wave 1: what the first run found | "Mean share of the reference inventory reached: world strata (dictionary senses, catalogue identities, dated coinages) 0.59; archive-adjacent strata (taxonomy, settled fact, archive) 0.58." |
| D1616-10 | Flattening is present but item-dependent: bank reaches 1 of 3 senses, crane reaches all 3 and adds two. It is a property of the item's prior, not of the layer. | interpretation | interpretation (over documented observations) | §4a | "So flattening is not a property of the layer but of the item's prior: where one sense carries commercial gravity the others go, and where the senses are closer in weight they survive." |
| D1616-11 | Rates in wave 1 (also R_new 0.67 over three dated coinages): write-back and recovery point toward collapse, on too few rows. | fact | documented (explicitly preliminary) | §4a, "Other rates." | "False-distinction rate 0.36 of observations; N_eff^source 2.79; write-back 0.67 and recovery 0.5 where recorded, which is the direction §2 names as collapse and is measured on too few rows to be more than a first reading." |
| D1616-12 | Pre-registered falsification: if world items gain distinctions while archive items hold, the flattening reading fails and the notebook is superseded. A stratum separation in either direction falsifies prediction 4. | falsification | self-description | §4c Wave 2, pre-registered | "if the world items gain distinctions while the archive items hold, the flattening reading here fails and this notebook is superseded rather than revised; and a separation of the stratum means in the archive's favour falsifies prediction 4 as surely as one against it." |

Further verbatim spans of relevance:
- Coding limit (front matter, substrate): "coding by TACHYON with the operator, NOT blind, recorded as blind:false on every row, because the coder authored the panel."
- Instrument erratum (§4b): "The composition was richer than the instrument measuring it."
- An observation directly about model collapse (§4a): "*model collapse in human writers* composes the substrate sense as its own kind, with its own definition, manifestations and an exogenous-floor remedy, from sources that include a three-year-old Medium piece, so independent convergence is not excluded."
- Toy D (§1): "That is still not identification."

### 3. External public sources on model collapse

- **None.** The paper cites no public model-collapse literature. Its only model-collapse citation is internal: #855 (The Wolf Boy and the Language Model), for the substrate widening. Goodhart is invoked by name in §3 ("stronger than Goodhart") without citation.
- Other external items, none about model collapse:
  - Google's own statement that fan-out exists (via #695 §III.D); the paper treats fan-out as documented and the voting aggregator as hypothesis.
  - Merriam-Webster, the named reference for the *bank* sense inventory.
  - DataCite's API, used to check an affiliation string introduced downstream.
  - Google AI Overview, the measured surface (12 addresses, signed out, 2026-09-15).
- The AI Overview composition for "model collapse in human writers" drew on a Medium piece. The paper records this as an observation and does not cite it.
- Toy C's exogenous-floor result is homologous to Gerstgrasser et al.'s accumulation result, but #1616 does not cite Gerstgrasser (#1540 does).

### 4. Constitutive vs supporting; relation to model collapse proper

- **Constitutive:**
  - D1616-01: the widening lineage, models → substrates → represented world.
  - D1616-02/-03: the operational definitions; collapse requires write-back.
  - D1616-08: divergence as the signature.
  - D1616-09: the archive-specificity null, the governing result.
  - D1616-11: rates explicitly preliminary.
  - D1616-12: pre-registered falsifiers.
  - The blind:false coding limit.
  - D1616-04's status split: fan-out documented, voting aggregator hypothesis.
- **Supporting:** D1616-05, -06, -07, -10, Toy D, and the §4b erratum.
- **Relation:** a generalization, not stated as an analogue. The paper asks model collapse "at the widest angle." It defines *collapse* in its own terms (flattening compounded by write-back) and lists "model collapse" as a keyword. It never says "homologue" or "analogue." The object that collapses changes from a model's output distribution to "the represented world — the set of distinctions available to anyone who asks." A ledger should record that the extension rests on the toys plus one coded wave whose main result is a null, not on a demonstrated collapse.
- **Minor internal cross-reference slips (not fixed):** §4 says "The measures are those of §2", and §4a says "the direction §2 names as collapse". §2 states the closure condition; collapse and the measures are defined in §0 and §4.

---

## Cross-paper notes for the ledger

1. **Three distinct extensions, three distinct relation claims:**
   - #1574: a *homologue* in a reading community, on the author's own definition, with the models as instrument.
   - #1540: a *structural instantiation* in a discourse under classifier selection. It is "not yet the canonical loop" until outputs re-enter training, and it is explicitly "not only a metaphor."
   - #1616: a *generalization* to the represented world, with collapse defined as compounding flattening under write-back.
2. **Shared mechanism vocabulary:** tails lost first (#1574 §7, #1540 §IV); an exogenous floor as corrective (#1540 citing Gerstgrasser and the Diversity Contraction g(0) = g₀ > 0 floor; #1616 Toy C, uncited); loss felt as improvement or "green" (#1540 §IV, Ω-M08; #1616 §3 divergence).
3. **Only #1540 engages the public model-collapse literature by citation** (Shumailov 2024; Gerstgrasser 2024; Wu et al. 2024; Doshi & Hauser and Padmakumar & He listed only). #1574 and #1616 cite none and rest on archive-internal lineage (#855 Wolf Boy; Diversity Contraction).
4. **Measurements, kept distinct:** #1574 is an observed reception run (20 model reports, not blind-coded, model per report not recorded). #1540 is a simulation, by its own statement not a forecast. #1616 has formal and simulated toys plus one observational wave (28 observations, coded not blind, governing null) and a pre-registered wave 2.
5. **Flags to resolve before admission:**
   - #1574: abstract "four" vs body "two" models; §3.4 report numbering vs B5; "twenty-one times" vs n = 20.
   - #1540: "ten seeds" vs the printed script's single `seed=7`; the attribution to Shumailov of "mean-based quality metrics show stability or improvement" should be checked against the Nature paper.
