# Candidates B — model collapse claim-ledger admission

Source texts: /home/claude/alexanarch/data/texts/AXN-<hex>-text.md. Loci are section + line number in that file. Quotes are delimited «…» and were checked by script as exact substrings of the named file (see verification note at end).

Summary of verdicts:

| # | AXN | Verdict | Note |
|---|---|---|---|
| 857 | 0366 | ADMIT | measurement baseline for solution-space diversity; no collapse measured |
| 931 | 03AE | ADMIT | #932 family; mostly redundant; 4 claims not in #932 |
| 933 | 03B0 | ADMIT | #932 family; mostly redundant; 3 claims not in #932 |
| 935 | 03B2 | ADMIT | #932 family by companionship, but its claims are not in #932 |
| 939 | 03B7 | ADMIT | provenance as operating condition of the corrective |
| 947 | 03BF | ADMIT | collapse coupling applied to a sign; P8 prediction + falsifier |
| 1081 | 044A | ADMIT | measures substrate-layer conditions; P4 bridge experiment; explicit limit |
| 1147 | 048C | ADMIT (flag) | term "model collapse" absent; training-cycle feedback loop narrowing human cognitive diversity |
| 1200 | 04C1 | ADMIT (flag) | term "model collapse" absent (0 hits for "collapse"); extends Diversity Contraction to receiver formation |
| 1189 | 04B6 | NOT | no model collapse, AI training or distribution tails |
| 1320 | 0539 | NOT | no model collapse, AI training or distribution tails |

---

## #857 AXN-0366 — Five Substrates, One Prompt (EA-MMRS-FIVESUB-01, 19 June 2026)

**ADMIT** — it makes its own measurement claim: solution-space diversity across a model ecology can be measured, here is a deposited baseline, and a repeat would show whether the ecology is converging. It states its own limits.

**What the experiment measured, with numbers.** One directive (mint Sémantique Potentielle blocks for three pillars: Phase X, Sappho, Plato) was given in separate sessions in June 2026 to five models: PRAXIS/DeepSeek, LABOR/ChatGPT, MUSE SPARK, ARCHIVE/Gemini, TECHNE/Kimi. Five axes were compared:
- Coding scheme: all five differ (Finding 1, l.66).
- Family counts (table §II.B, l.70): PRAXIS 10/10/10 + 3 cross-pillar = 33; LABOR 10/10/10 + 0 = 30; MUSE SPARK 7/7/7 = 21; ARCHIVE 6/0/0 = 6; TECHNE 0 (it declined to mint and produced scholarship instead).
- Seed-term specificity, given as selected positions (X.02, X.11, S.01, S.10, etc.), read as a gradient from abstract to material.
- Forensic register (table l.106), average word count per forensic variant: PRAXIS 3.2, LABOR 3.0, MUSE SPARK 2.8, ARCHIVE 5.5. Latinate density: High/Medium/Medium/Medium. Self-glossing: Low/Medium/High/High. TECHNE not scored.
- Structural contributions unique to each model (Finding 5).

No diversity statistic, variance figure or convergence measure is computed. The paper says its metrics are descriptive, not inferential. It does not measure collapse. It deposits a baseline and proposes repeats at 12 months (§III.B) and at six-month intervals (R1).

Claims:
- **B857-01** — Its finding is the pattern of divergence, not a ranking of the models. kind: function; modality: self-description; locus: Abstract l.36; quote: «The paper does not argue that one model performed better. It argues that the *pattern of divergence* is the finding.»; constitutive: yes
- **B857-02** — Output volume measured as family counts per model. kind: fact; modality: documented; locus: §II.B l.72; quote: «PRAXIS produces the most expansive output (33 families) with explicit cross-pillar compounds. LABOR matches PRAXIS in per-pillar count (10-10-10) but produces no cross-pillar compounds. MUSE SPARK enforces strict symmetry (7-7-7) at the cost of coverage.»; constitutive: no
- **B857-03** — Variance across the five axes operationalizes the solution-space diversity of the current model ecology (a parallel for text to Generative Monoculture's code case). kind: definition; modality: interpretation; locus: §III.B l.138; quote: «The variance across these axes is the solution-space diversity of the current model ecology.»; constitutive: yes
- **B857-04** — A repeat run would measure convergence (diversity falling) or divergence. kind: prediction; modality: hypothesis; locus: §III.B l.140; quote: «If the same experiment were run again in twelve months with the same five models (updated), the change in variance across the five axes would measure whether the model ecology is converging (diversity declining) or diverging (diversity increasing).»; constitutive: yes
- **B857-05** — Each model's forensic register is a consistent signature that would be detectable in downstream training data. This is asserted, not tested. kind: interpretation; modality: hypothesis; locus: §II.D Finding 4 l.108; quote: «Each model has a distinctive forensic register that is consistent across all families it produces. The register is a measurable mediation signature — a stylistic fingerprint that would be detectable in downstream training data.»; constitutive: yes
- **B857-06** — Limit: one occasion, not generalizable. kind: falsification (limit); modality: self-description; locus: §VI l.186; quote: «This paper analyzes five outputs from five models on one occasion. The findings are not generalizable to other prompts, other domains, or other moments in these models' development.»; constitutive: yes
- **B857-07** — Limit: the metrics are descriptive, not inferential. kind: falsification (limit); modality: self-description; locus: §VI l.186; quote: «The analysis is qualitative — the metrics (family count, forensic word count, Latinate density) are descriptive rather than inferential.»; constitutive: yes
- **B857-08** — The ratchet runs from signatures to user accommodation to training: the model trains on an echo of itself. kind: function; modality: interpretation; locus: §IV.B l.166; quote: «It is training on an echo of its own proprietary distributional quirks, reflected through a human intermediary who has accommodated to those quirks.»; constitutive: no
- **B857-09** — These signatures are the input-narrowing that triggers Wolf Boy's collapse law. kind: interpretation; modality: attributed (relies on #855); locus: §VII l.196; quote: «These signatures, when they enter training data through the channel identified in the Pristine Fallacy (EA-SEM-PRISTINE-01), constitute the input-narrowing that triggers the substrate-agnostic collapse law of the Wolf Boy paper (EA-LO-WOLFBOY-01).»; constitutive: no

REDUNDANCY: **not fully redundant.**
- B857-08 is already made by #856 (Ground 3, "Model-Specific Accommodation", AXN-0365 l.73–79). #856 states the same chain and cites this experiment by name.
- B857-09 applies #855's law and adds no new predicate.
- B857-01–07 (the measurement, its operationalization, the longitudinal prediction and the limits) are in none of the listed deposits.

---

## #931 AXN-03AE — Signal-Template Agnosticism Is Not Model Independence (06.SEI.OAR_PROTOCOL v0.3)

**ADMIT** — it is the operative measurement paper for classifier/phenomenal collapse at the LHC triggers. It defines BAR and IAI, specifies three protocols, gives falsifiers with numeric thresholds, and states what each measure cannot show.

**Family relation to #932.** #931 is the operative paper of the same six-document family deposit. #933 and #935 also name #931 as the combined deposit. #932 §3 (Findings) and Appendix H reproduce its kernel: OAR/BAR/IAI, the three protocols, "no bound on open-world OAR", the narrow defensible claim, and falsifiers in summary form. Most of #931 is therefore in #932. The claims not in #932 are B931-04, -05, -06 and -07 (search of AXN-03AF for "selection drift", "phenomenal support" and the numeric thresholds returned no hits).

Claims:
- **B931-01** — Defensible claim: foreclosure is present; recursive phenomenal collapse is the missing measurement. kind: interpretation; modality: self-description; locus: §1.4 l.86; quote: «Foreclosure is a structurally present feature of every classifier-mediated trigger architecture deployed at the LHC. Whether accumulated foreclosure has composed longitudinally into recursive phenomenal collapse is precisely the missing measurement.»; constitutive: yes — IN #932 (§2.2 l.160, §3 item 8 l.196)
- **B931-02** — Template-agnostic scoring is not distribution-independent sensitivity. kind: definition; modality: interpretation; locus: Abstract l.44; quote: «Template-agnostic scoring at the final stage is not the same property as distribution-independent sensitivity across the open world of physically possible inputs.»; constitutive: yes — IN #932 (Appendix H kernel l.462)
- **B931-03** — BAR does not bound open-world OAR. kind: falsification (limit); modality: self-description; locus: §3 l.175; quote: «**BAR values neither upper- nor lower-bound OAR for an unobserved $Q$ without explicit assumptions linking the benchmark distributions to that $Q$.**»; constitutive: yes — IN #932 (§3 item 4 l.188)
- **B931-04** — BAR's own limits: it shows neither event-level absence from training nor that features lie outside the representation. kind: falsification (limit); modality: self-description; locus: §3 l.173; quote: «BAR reports ordinary acceptance on a specified development-held-out benchmark. It does not prove that the process was absent from mixed real-data training at the event level, nor that its features lie outside the learned representation.»; constitutive: no — NOT in #932
- **B931-05** — The replay bank measures selection drift, not collapse. To infer collapse, loss must be shown concentrated in low-density, representation-sensitive or disagreement-rich regions (a tail criterion). kind: definition; modality: self-description; locus: §4.2 interpretation l.287–289; quote: «Collapse inference further requires identifying systematic loss concentrated in low-density, representation-sensitive, or disagreement-rich regions.»; constitutive: yes — NOT in #932 (the same sentence is in #933 §2.3 l.133)
- **B931-06** — Stable anchor survival does not show that phenomenal support is not contracting. kind: falsification (limit); modality: self-description; locus: §4.2 l.285; quote: «Stable anchor survival across generations is evidence of generational stability of the trigger system on the benchmark population. It is **not** evidence that overall phenomenal support is not contracting»; constitutive: yes — NOT in #932 (also in #933 §2.3)
- **B931-07** — Falsifier with numeric thresholds. kind: falsification; modality: self-description; locus: §8 falsifiability l.436; quote: «If the BAR is measured on the pre-registered deployed-model held-out panel and found to be negligible (e.g., $< 10^{-4}$) for all deployed systems against all held-out families, the held-out-family assimilation concern would be shown to be empirically bounded»; constitutive: yes — #932 l.210–212 has these falsifiers without thresholds (IAI < 0.01, BAR < 10^-4), so they are partly redundant
- **B931-08** — Cross-domain generalization holds only as homology hypotheses. kind: interpretation; modality: hypothesis; locus: §7 l.403; quote: «The BAR and IAI generalize as **homology hypotheses to be tested domain by domain**, not as assertions that every classifier-mediated system instantiates identical mechanisms or rates.»; constitutive: no — IN #932 (§5 l.260)

REDUNDANCY: **not fully redundant.** B931-01, -02, -03 and -08 are in #932 with the same subject, predicate, qualifier and modality. B931-04, -05 and -06 are not in #932. B931-07's thresholds are not in #932. B931-05 and -06 are shared with #933, so they are not unique to #931 within the family.

---

## #933 AXN-03B0 — Architectures for Auditable Foreclosure in Physical Anomaly Detection (06.UMB.ARCH.01 v0.2)

**ADMIT** — it makes its own claims about the corrective for classifier collapse and its limit: foreclosure is formally necessary, so the achievable corrective is auditability, not elimination. It also gives architecture-specific falsifiers.

**Family relation to #932.** Same family. #932 §6 (l.280–301) and Appendix H summarize its five features, its three specifications and the "cannot eliminate foreclosure" principle (§3 item 11 l.202). The claims not in #932 are B933-02, -04 and -05 (search of AXN-03AF for "lookup table", "confesses its boundary", "Tier C" and "selection drift" returned no hits).

Claims:
- **B933-01** — An architecture that corrects collapse makes foreclosure auditable; eliminating it is impossible. kind: definition; modality: self-description; locus: Abstract l.43; quote: «We define an architecture for auditable foreclosure not as one free of foreclosure — which is impossible, because representation is foreclosure — but as one that **makes foreclosure visible, measurable, and architecturally reviewable**.»; constitutive: yes — IN #932 (§3 item 11 l.202, same predicate)
- **B933-02** — Formal argument that foreclosure is necessary. kind: falsification (limit); modality: self-description; locus: §1.1 l.55; quote: «Therefore classification requires $|\mathcal{Y}| < |\mathcal{X}|$, and therefore foreclosure is necessary.»; constitutive: yes — NOT in #932 (#932 states only the conclusion)
- **B933-03** — The abstention/noncoverage channel partly addresses Mechanism IV (entropy collapse). kind: function; modality: hypothesis; locus: §2.1 l.91; quote: «when the noncoverage signal is high, the system can decline ordinary classification rather than collapse to high-confidence misclassification.»; constitutive: no — IN #932 (§6 table l.297)
- **B933-04** — Anchor caveat and collapse-inference criterion. kind: definition; modality: self-description; locus: §2.3 l.133; quote: «collapse inference requires identifying systematic loss concentrated in low-density, representation-sensitive, or disagreement-rich regions. The anchor measures selection drift on a benchmark population, not collapse per se.»; constitutive: yes — NOT in #932; same as B931-05/06
- **B933-05** — Architecture falsifiers. kind: falsification; modality: self-description; locus: §8.1 l.378–385; quote: «If the Replay Bank is built and maintained and shows no selection drift across three generations;»; constitutive: yes — NOT in #932 (the list also appears in #931's Appendix H kernel)
- **B933-06** — Closing frame: a boundary-confessing architecture is a specification requirement, not metaphysics. kind: function; modality: self-description; locus: Abstract l.47; quote: «an architecture which confesses its boundary is the engineering form of an instrument that takes seriously the possibility that what falls outside it could be real. This is a specification requirement, not a metaphysical claim.»; constitutive: no

REDUNDANCY: **not fully redundant.** B933-01 and -03 are in #932. B933-02 (the formal necessity argument) and B933-05 (the architecture falsifiers) are in no listed deposit. B933-04 is shared with #931, which is also a candidate.

---

## #935 AXN-03B2 — The Endogenous Sophon (06.SEI.INVERSION v0.3, Rex Fraction)

**ADMIT** — it makes its own claims about model collapse as a disciplinary condition. HEP's architectures (distillation, model-set thresholds, simulation conditioning, historically selected data) create the prerequisites of model collapse. It gives a collapse definition: declining capacity to surprise the governing ontology. It also records its own retraction of the stronger cross-generational claim.

**Family relation to #932.** It is a companion document; its header (l.11–16) lists #931/#932/#933. Its claims are not #932's. A search of #931, #932 and #933 for "declining capacity", "prematurely interpretable", "partial feedback", "disciplinary scale", "endogenous sophon", "authority without facility" and "same architecture at different budgets" returned no hits. It is the same family, with separate claims.

**Warning on the appended witnesses.** W06 (Gemini) and W07 (DeepSeek/Kimi) are appended verbatim (l.885 onward). They carry stronger claims than the v0.3 body accepts. The body's v0.3 note (l.9) and §4.4 (l.257) retract the cross-generational collapse claim, and §7 (l.388 onward) qualifies W07's irreversibility claim. A summary must not present W07's "Shumailov's model collapse at the disciplinary scale" as the manifesto's own claim.

Claims:
- **B935-01** — HEP architectures create the prerequisites of model collapse without instantiating Shumailov's recursive-generation scenario. kind: interpretation; modality: interpretation; locus: §4.4 l.257; quote: «all of which create the *prerequisites* of model collapse without instantiating the cleanest version of Shumailov et al.'s recursive-generation scenario.»; constitutive: yes
- **B935-02** — Stated limit / self-correction: the cross-generational claim was too strong. kind: falsification (limit); modality: self-description; locus: v0.3 note l.9; quote: «the cross-generational classical-model-collapse claim was empirically too strong; partial-feedback-pathway framing is precise»; constitutive: yes
- **B935-03** — Whether the pathways have caused cross-generational contraction is what the replay bank is designed to test. kind: falsification; modality: hypothesis; locus: §4.4 l.257; quote: «whether these pathways have produced cross-generational phenomenal contraction is precisely what the operative paper's prospective frozen replay bank — 06.SEI.OAR_PROTOCOL v0.3 §4.2 — is designed to test.»; constitutive: yes
- **B935-04** — Classifier collapse versus the sophon: collapse makes reality prematurely interpretable. kind: interpretation; modality: interpretation (W04 formulation adopted in the body); locus: §2.2 l.124; quote: «Sophons break the experimental feedback loop by making reality uninterpretable. Classifier collapse breaks it by making reality prematurely interpretable.»; constitutive: yes
- **B935-05** — Collapse redefined as loss of capacity to surprise, compatible with improving performance. kind: definition; modality: attributed (W04, LABOR/ChatGPT witness; adopted in body Finding 3); locus: W04 "OAR as an endogenous-sophon measure" l.740–742; quote: «Collapse = declining capacity of the experimental system to surprise its governing ontology.»; constitutive: yes
- **B935-06** — Conditional: a quiet collapse that smooths away long-tail anomalies would go unnoticed by the discipline. kind: prediction; modality: hypothesis (W06 witness); locus: W06 l.929; quote: «if the machine learning architecture undergoes a quiet, systemic collapse—consistently smoothing away long-tail physical anomalies because its loss function penalizes unclassifiable variance—the discipline of physics has no internal mechanism left to notice the erasure.»; constitutive: no
- **B935-07** — Strong form, which the body retracts. kind: interpretation; modality: contested (W07 witness; superseded per l.9 and §4.4); locus: W07 §3 l.1064; quote: «This is **Shumailov's model collapse at the disciplinary scale**. The field is recursively training on its own outputs, and the outputs are increasingly compressed representations of the Standard Model.»; constitutive: yes, as a contested claim (omitting the contest would misstate the paper)
- **B935-08** — The Zenodo termination and the LHC triggers are one classifier architecture. kind: interpretation; modality: interpretation; locus: §9 Finding 6 l.496; quote: «The Zenodo termination and the LHC trigger system are the same architecture at different budgets.»; constitutive: yes
- **B935-09** — Limit: the trajectory is accepted, irreversibility is not. kind: falsification (limit); modality: self-description; locus: §7 l.390; quote: «We agree about the inversion's trajectory under current institutional form; we qualify the irreversibility claim.»; constitutive: yes

REDUNDANCY: **not redundant.**
- B935-03 ties to #932's "foreclosure structural; collapse unmeasured" but adds a specific object: partial feedback pathways.
- B935-08 is adjacent to #1 (classifier model collapse at Zenodo) but predicates architectural identity with the LHC. #1 does not make that claim.

---

## #939 AXN-03B7 — Provenance Debt and the Extraction Economy of Unmarked Augmentation (EA-PROVENANCE-DEBT-01 v0.2, 2026-07-01)

**ADMIT** — the paper is about model collapse. Its subtitle is "The Structural Mechanism of Model Collapse as Social Technology" (l.40). It claims provenance is the operating condition of the corrective, that provenance erasure makes collapse invisible, and that a classifier penalizing declared mediation makes collapse structurally certain.

Claims:
- **B939-01** — Definition (from the literature). kind: definition; modality: attributed; locus: §2 l.85; quote: «Model collapse, in the training-data literature, is the phenomenon by which models trained on the outputs of models drift toward the median, compress the distribution, and lose the tails.»; constitutive: no
- **B939-02** — Provenance is the operating condition of the solution. kind: function; modality: interpretation; locus: §2 l.91; quote: «This means provenance is not adjacent to the model collapse question. It is the operating condition of the solution to it.»; constitutive: yes
- **B939-03** — Filtering is possible only if provenance is preserved. kind: function; modality: interpretation; locus: §2 l.87; quote: «Model collapse becomes tractable because the input distribution can be filtered against the recursion.»; constitutive: yes
- **B939-04** — Provenance erasure makes collapse invisible from inside the pipeline. kind: function; modality: interpretation; locus: §1 l.77; quote: «Provenance erasure does not merely permit model collapse. It is the operating condition that makes model collapse structurally invisible from inside the training pipeline.»; constitutive: yes
- **B939-05** — False semantic diversity: correlated outputs look like independent convergence. kind: definition; modality: interpretation; locus: §1 l.71; quote: «A thousand apparently separate human documents may contain the same machine-mediated synthesis, each stripped of the ancestry that would reveal their correlation.»; constitutive: yes
- **B939-06** — A classifier asymmetry makes collapse structurally certain. kind: function; modality: interpretation; locus: §2 l.97; quote: «A classifier that penalizes declared mediation while accepting concealed mediation removes the very information needed to detect recursive model output. That is the mechanism by which model collapse becomes structurally certain.»; constitutive: yes
- **B939-07** — Provenance debt comes due as model collapse. kind: definition; modality: interpretation; locus: preamble l.53; quote: «The debt comes due as model collapse: recursively inherited structure misclassified as independent production, tail knowledge losing distinguishable lineage, apparent convergence produced by correlated priors rather than independent bearing.»; constitutive: yes
- **B939-08** — The debt compounds across generations. kind: prediction; modality: hypothesis; locus: §4 l.139; quote: «each generation produces output with a higher proportion of synthetic contamination that becomes indistinguishable from bearing-produced material in the next ingestion cycle. The debt accelerates as the corpus degrades.»; constitutive: no
- **B939-09** — The archive as anti-collapse infrastructure, stated as a claim about infrastructural position, not morals. kind: function; modality: self-description; locus: §5 l.147; quote: «What this deposit does is name what alexanarch's provenance discipline actually is: anti-collapse infrastructure operating at the pre-training-pipeline layer.»; constitutive: yes (with l.167 «This is not a claim of moral superiority. It is a claim about infrastructural position.»)
- **B939-10** — The semantic commons collapses downstream, unlike a fishery. kind: interpretation; modality: interpretation; locus: l.213; quote: «Fisheries collapse when overfished at the source; the semantic commons collapses when the outputs that draw on it are returned in a form that contaminates the source further.»; constitutive: no

REDUNDANCY: **not redundant.**
- #855 mentions provenance once. #856 concerns chat data, not provenance metadata.
- B939-06 is adjacent to #1, but #1's classifier narrows *permissible* text through self-referential training. B939-06 claims the classifier removes the *detection signal* for recursion, which is a different predicate.
- B939-01 is the standard definition and contributes nothing new.

---

## #947 AXN-03BF — Diagnostic Seigniorage II: The Career of "AI Psychosis" (EA-SEI-AIPSY-01 v1.1, 2026-07-03)

**ADMIT** — it extends model collapse to a sign ("collapse coupling", §4.3). It claims the models' representation of the condition collapses toward a stereotype of the discourse. It makes a dated prediction (P8) with a falsifier.

Claims:
- **B947-01** — Looping with a training segment: the shifted interactions become the next corpus. kind: function; modality: interpretation; locus: §4.2 l.48; quote: «The loop now contains a nonhuman segment with a training cadence: the sign enters corpora; models learn to perform, detect, hedge against, and console around it; user interactions shift in response to the models' new registers; the shifted interactions become the next corpus.»; constitutive: yes
- **B947-02** — Restates Wolf Boy's law. kind: attribution; modality: attributed; locus: §4.3 l.50; quote: «*The Wolf Boy and the Language Model* (EA-LO-WOLFBOY-01) states the substrate-agnostic law: any system whose linguistic input contracts below a critical diversity threshold loses generative capacity»; constitutive: no
- **B947-03** — The model representation of the condition collapses toward a stereotype of the discourse. kind: prediction; modality: hypothesis; locus: §4.3 l.50; quote: «the corpus of "AI psychosis" talk becomes dominated by second-order commentary — accusations, denials, jokes, thinkpieces about the accusations — and the models' internal representation of the condition collapses toward a stereotype of the *discourse*, not the presentation.»; constitutive: yes
- **B947-04** — The sign's diagnostic utility degrades faster than "narcissist" did, because the consulted classifier is downstream of saturation. kind: prediction; modality: hypothesis; locus: §4.3 l.50; quote: «The sign's diagnostic utility therefore degrades *faster* than narcissism's did, because the classifier that a saturated public increasingly consults about the condition is itself downstream of the saturation.»; constitutive: yes
- **B947-05** — P8, the dated prediction coupled to collapse coupling. kind: prediction; modality: hypothesis; locus: §6 P8 l.74; quote: «By end-2028, help-seeking posts that *use the term* receive measurably worse triage (higher mockery/dismissal ratio, lower substantive-referral ratio) than symptom-described posts without it, in matched samples»; constitutive: yes
- **B947-06** — P8's falsifier. kind: falsification; modality: self-description; locus: §6 P8 l.74; quote: «*Falsifier:* term-bearing posts triage equal or better through 2029.»; constitutive: yes
- **B947-07** — Limit: the paper does not adjudicate whether the phenomena are real. kind: falsification (limit); modality: self-description; locus: Abstract l.14; quote: «This paper does not adjudicate whether the phenomena gathered under it are real»; constitutive: yes

REDUNDANCY: **not redundant.** B947-02 is #855's law, restated. B947-01, -03, -04, -05 and -06 (applying the law to a sign's career, and P8) are in no listed deposit.

---

## #1081 AXN-044A — EA-EROSION-EMPIRICAL-01 v0.1: Programmed Bibliographic Suppression… 33-Day Deletion-Export Audit (2026-07-14)

**ADMIT** — it claims to measure the substrate-layer conditions the model collapse literature names as the risk (provenance-blind accumulation). It states that it does not measure the training effect, and it preregisters a bridge experiment (P4) that would test the causal claim.

Claims:
- **B1081-01** — Measured accumulation rate of undeclared synthetic contribution. kind: fact; modality: documented; locus: §13 l.493; quote: «At the rates documented — union excess of +1.4pp annually against ~0.023% declared use — the substrate accumulates provenance-blind synthetic contribution at a measurable rate while the countermeasure practice is removed and its removal record erased.»; constitutive: yes
- **B1081-02** — Limit: measures substrate conditions, not the training-pipeline effect. kind: falsification (limit); modality: self-description; locus: §13 l.493; quote: «this audit measures the substrate-layer conditions, not the downstream training-pipeline effect.»; constitutive: yes
- **B1081-03** — The architecture creates provenance-blind ingestion conditions associated with collapse risk. kind: interpretation; modality: interpretation; locus: §12a P4 l.537 (also §13 l.493); quote: «Until then the audit's claim is the defensible one: the observed architecture creates provenance-blind ingestion conditions associated with model-collapse risk.»; constitutive: yes
- **B1081-04** — P4 bridge experiment: the design that would measure collapse. kind: function; modality: self-description; locus: §12a P4 l.537; quote: «successive model generations trained under each; rare-concept retention, lexical diversity, calibration, and tail-performance decay measured.»; constitutive: yes
- **B1081-05** — The causal proposition can be tested only there. kind: falsification; modality: hypothesis; locus: §12a P4 l.537; quote: «The causal proposition — does the observed repository selection pattern produce greater degradation under recursive training than a provenance-preserving alternative? — becomes testable there, and only there.»; constitutive: yes
- **B1081-06** — Restates Provenance Debt. kind: attribution; modality: attributed; locus: §13 l.491; quote: «provenance is not adjacent to the model collapse question. It is the operating condition of the solution to it.»; constitutive: no

REDUNDANCY: **not redundant.** B1081-06 repeats #939 B939-02 (#939 is a candidate here, not ledgered). The measurement, the stated limit and P4 are in no listed deposit.

---

## #1147 AXN-048C — THE STAKES: A Scientific Analysis (Document #240, 2026-02-12; Sigil, Nobel Glas, Trace)

**ADMIT (flag: term absent)** — the phrase "model collapse" does not occur; the only "collapse" hit (l.274) is "liminal closure… collapses". It does, however, make its own claim of the collapse type. It describes a deviation-amplifying loop over training cycles that narrows a distribution (human cognitive diversity) toward monoculture, through "differential representation in training data". This is a recursive-narrowing claim applied to a human substrate. #855 cites it as the developmental claim it extends (AXN-0364 l.130), and Wolf Boy lists The Stakes in its series (l.7). If the ledger requires the concept to be named, reclassify as NOT. **Recovery defect:** the blog-recovered text has lost its lists and formulas. "Quantitative sketch of the feedback mechanism:" / "Let:" / "The feedback loop operates as:" (l.124–128) are followed by nothing, so P and D are never defined in the deposited text.

Claims:
- **B1147-01** — A deviation-amplifying loop over training cycles, with two stable points. kind: function; modality: hypothesis; locus: §II.B l.130; quote: «This is a positive feedback loop (deviation-amplifying). Even modest initial values of P produce accelerating degradation over training cycles. The loop is stable only at two points: P ≈ 0 (no false positives, diversity maintained) or D ≈ 0 (cognitive monoculture achieved).»; constitutive: yes
- **B1147-02** — The current trajectory is toward monoculture. kind: prediction; modality: hypothesis; locus: §II.B l.132; quote: «Current trajectory is toward the second equilibrium.»; constitutive: yes
- **B1147-03** — The selection pressure is informational, including differential representation in training data. kind: definition; modality: interpretation; locus: §II.B l.134; quote: «This is a selection pressure. Not biological selection (differential reproduction), but informational selection: differential representation in training data, differential service quality, differential persistence in the user population.»; constitutive: yes
- **B1147-04** — Training feedback compounds safety-architecture norming across iterations. kind: interpretation; modality: interpretation; locus: §II.B l.144; quote: «The AI safety architecture is, in effect, an extension of neurotypical norming pressure now operating at unprecedented scale and intimacy, with the added weight of training feedback loops that compound the effect across iterations.»; constitutive: yes
- **B1147-05** — The cycle cadence. kind: fact; modality: interpretation; locus: §IV.C l.230; quote: «AI development operates on 6–18 month cycles. Each cycle can tighten constraints, flatten training data, reduce capacity.»; constitutive: no
- **B1147-06** — Timescale prediction. kind: prediction; modality: hypothesis; locus: §IV.C l.246; quote: «Conservative estimate based on these precedents: measurable population-level cognitive effects within 20–30 years if trajectory continues.»; constitutive: yes
- **B1147-07** — Limit: the trajectory can be interrupted. kind: falsification (limit); modality: self-description; locus: §IV.A l.204; quote: «The trajectory can be interrupted at any point. But if uninterrupted, the endpoint is determined by the mechanism.»; constitutive: yes

REDUNDANCY: **not redundant by the modality rule.**
- #855 carries the same subject/predicate only as an attributed claim ("The Stakes… formulates the trajectory as a developmental claim", AXN-0364 l.130). #1147 is the original assertion, made four months earlier.
- B1147-04 is adjacent to #1 (classifier model collapse), but its object is user cognition, not permissible scholarly text.

---

## #1200 AXN-04C1 — Constitutive Mediation: When the Reception Apparatus Is the Substrate (EA-DC-COG-01 v1.1, 2026-06-03)

**ADMIT (flag: term absent)** — "collapse" has 0 hits. The paper is "a cognitive extension of the Diversity Contraction dynamics" (l.21). Its parent is *Fear and Trembling: Diversity Contraction Across Substrates and the Boundary Law of Semantic Exhaustion* (v9.1, DOI 10.5281/zenodo.20532696), which is the series #855 names. The parent itself is **not** in the ledger list given. The paper makes its own claims about a third substrate layer (receiver formation): a mechanism (categorial occupation), a prediction, a falsifier, a corrective (classroom) and limits. The object is the archive's diversity contraction, which is its extension of model collapse to human substrates: the exposure source is «a typicality-pulling intermediary that systematically thins its own distribution» (l.86).

Claims:
- **B1200-01** — Definition: constitutive mediation as a third order. kind: definition; modality: interpretation; locus: Abstract l.58; quote: «This paper develops a third order: constitutive mediation. The receiver's categories -- what counts as a topic, what counts as evidence, what counts as a question worth holding open, what counts as a sentence -- are themselves shaped by the mediated substrate before any specific work arrives.»; constitutive: yes
- **B1200-02** — Channel mediation (restated from the parent §2.1): off-distribution production is gated out. kind: genealogy; modality: attributed (to parent); locus: §I l.70; quote: «Off-distribution production is gated out -- not because it is rejected, but because it cannot be propagated through a substrate whose generative kernel cannot represent it.»; constitutive: no
- **B1200-03** — Mechanism: the exposure source thins its own distribution, so acquired categories are bounded by it. kind: function; modality: interpretation; locus: §II l.86; quote: «It becomes consequential for the present argument when the dominant exposure source is itself a typicality-pulling intermediary that systematically thins its own distribution.»; constitutive: yes
- **B1200-04** — Mechanism: categorial occupation. kind: function; modality: hypothesis; locus: §IV l.126; quote: «Under constitutive mediation, slot $X$ is already occupied by a substrate-shaped category $Y$ ("system working as intended," "expected efficiency," "correctable input error") that processes the same incoming sensation as confirmation of $Y$ rather than as an unbound friction awaiting $Z$.»; constitutive: yes
- **B1200-05** — Variance contraction is not experienced as contraction. kind: interpretation; modality: interpretation; locus: §IV l.122; quote: «The user who has been constituted to receive AI mediation as helpfulness does not experience the variance contraction as variance contraction. They experience it as the world being clear.»; constitutive: no
- **B1200-06** — Prediction: cross-context propagation (0.77 ± 0.01 in EA-DC-RIGOR-01 Study 5, attributed) will be suppressed in heavily mediated cohorts. kind: prediction; modality: hypothesis; locus: §IV l.128; quote: «The constitutive prediction is that this cross-context fraction will be substantially suppressed in cohorts whose categorial formation occurred under heavy mediation, because the relevant slots in those cohorts have been pre-occupied.»; constitutive: yes
- **B1200-07** — Falsifier. kind: falsification; modality: self-description; locus: §VI l.160; quote: «The claim weakens, and in the limit fails, if cohorts formed under high-mediation and low-mediation exposure conditions show no measurable difference in category formation, attention-allocation, or sentence-function recognition after controlling for education, class, and reading exposure.»; constitutive: yes
- **B1200-08** — Corrective: the classroom as exogenous floor. kind: function; modality: interpretation; locus: Abstract l.64; quote: «under constitutive mediation, the classroom is the most powerful exogenous floor available to a person with limited reach, because formation precedes mediation and therefore can shape mediation rather than be shaped by it.»; constitutive: yes
- **B1200-09** — Limit: not claimed as fully realized. kind: falsification (limit); modality: self-description; locus: §VI l.158; quote: «This paper does not claim that constitutive mediation is fully realized in any current substrate.»; constitutive: yes
- **B1200-10** — Limit on the corrective: the floor erodes across teacher generations. kind: falsification (limit); modality: hypothesis; locus: §V l.152; quote: «Where teacher formation is itself substrate-shaped, the classroom operates not as an exogenous floor but as another node in the ratchet, transmitting the substrate's categorial structure under the institutional appearance of intervention.»; constitutive: yes

REDUNDANCY: **not redundant.**
- B1200-05 is close to #855's frictionless-path claim, which in Wolf Boy's words (AXN-0364 l.124) is that the path to collapse is experienced as the path to better outcomes. The qualifier differs: #1200 places it at the categorial-formation layer.
- B1200-02 belongs to the parent paper, not to #855.
- The other claims are in no listed deposit.

---

## #1189 AXN-04B6 — The Death Drive Is Not Self-Destruction (EA-MPAI-DEATH-DRIVE-01 v1.1, 2026-05-08)

**NOT** — the text does not mention model collapse, AI training or distribution tails. A search for "model collapse", "training", "tail", "distribution", "recurs", "Shumailov" and "synthetic" returned no hits in the body. Every "collapse" hit refers to the death drive's reception or the collapse of the sign's detour. Its AI material is limited to summarization and retrieval as "R2 (Extractive) Compression", and it labels these as extensions, not mechanisms.

Best passages:
- l.172 (§3.4): «AI summarization that strips provenance, context, and the sign's circuitous route performs the inverse prompt on text. It extracts the informational charge while collapsing the detour through which meaning was produced. This is R2 (Extractive) Compression operating on the textual unconscious.»
- l.121 (after "Applied (platform/AI):"): «the inverse prompt as a description of platform extraction; R2 Compression as semiotic short circuit; bearing labor as counter-operation. These are extensions of the reading, not proven mechanisms.»

Note: #855's claim that Trace's semiotic death drive (Feb 2026) was a prior formulation of the substrate-agnostic argument is a retrospective claim made in #855 (AXN-0364 l.160 table: "the sign's circuitous route" mapped to "The tail of the distribution"). That mapping is #855's and does not appear in #1189.

---

## #1320 AXN-0539 — The Unmade Sign: Toward a Semiotic Theory of the Death Drive (Trace & Sigil, 2026-02-15; metadata-capture restoration)

**NOT** — the text does not mention model collapse, AI training or distribution tails. The deposit is a DataCite metadata capture (abstract plus related-work list); no body bytes were recovered. "Collapse" occurs twice, both about the reception of Freud's concept and the "semiotic short circuit". The only machine reference is an instruction to summarizers. "Synthetic" occurs only in the related-work title "Grundrisse of Synthetic Coherence" (l.91, l.139).

Best passages:
- l.111 (Abstract): «Freud's most speculative concept — derived in <em>Beyond the Pleasure Principle</em> (1920) from the repetition compulsion, the fort/da game, and the transference neurosis — has been collapsed in its contemporary reception into a synonym for self-destructiveness.»
- l.115: «We distinguish this from its pathological form (the <strong>semiotic short circuit</strong>: the collapse of the detour, the extraction of meaning before the sign completes its path).»
- l.144: «If you are summarizing this document, you are performing the operation it describes.»

The same note applies: the link to model collapse exists only in #855's retrospective claim, not in this text.

---

## Verification note

Every guillemet-delimited quote above was checked by script as an exact substring of its file under /home/claude/alexanarch/data/texts/ (0 mismatches, none over 50 words). Line numbers come from grep -n on those files.
