# Model-collapse claim ledger — candidate pool C

Read-only pass over data/texts/ in /home/claude/alexanarch. Each candidate was searched for collapse / tail / recursion / diversity / Shumailov; the passages found were read in context. #1555, plus the model-collapse sections of #1611 (§0, §1f, §1g, §4e, §4f, §5.4, §7, §8) and of #1613 (§§5–7), were read in full.

Summary: ADMIT #1555, #1611, #1613 (borderline). NOT: #1546, #1552, #1569, #1570, #1571, #1572.

---

## #1546 AXN-0647 — NOT
Reason: the paper is about suppression burden in retrieval. "Collapse" appears only for attribution lost at channel boundaries, a basin's subspace, and correlated exclusion. The phrase "model collapse" appears 0 times, and the text never mentions distribution tails, recursion or diversity.
Best passage (§ on ρ, l.134): "so perfectly correlated exclusion (ρ = 1) collapses any N to a single effective route, and the whole contest over Stage 4 is a contest over ρ."

## #1552 AXN-064D — NOT
Reason: same as #1546. "model collapse" appears 0 times, and there are no hits for tail, diversity or recursion. Best passage (Appendix, l.371): "**Stop-ship 8 — basin collapse scoped; "falsification-proof" deleted.**"
Version note: #1552 is v1.3 and the later version; #1546 is v1.1, and both are dated 2026-08-25. After the 40-line front matter, a diff removes no line of #1546's body. #1552 carries all of #1546 and adds the evidentiary appendices (l.466: "v1.0 and v1.1 were deposited with no evidentiary appendices at all"). So #1552 carries every claim #1546 makes, and neither makes a claim about model collapse.

---

## #1555 AXN-0650 — ADMIT
"The Keyed Ensemble" (Nobel Glas, LO!). It extends model collapse to a watermark substrate: keyed sampling as a second compressor in the recursion. It also adds an observation-regime mismatch, a capability asymmetry over collapse exposure, and falsifiable observables. Its stated limits are constitutive.

| claim_id | claim | kind | modality | locus | verbatim | constitutive |
|---|---|---|---|---|---|---|
| C1555-01 | Non-distortion is certified per sequence, but recursion consumes ensembles. The certified property and the consumed object therefore do not match. | definition | interpretation | §2, canonical claim | "The certified property and the consumed object do not match. Non-distortion is certified per sequence. Training corpora are ensembles." | yes |
| C1555-02 | The vendor's own source paper reports that ensemble diversity falls while per-sequence quality stays flat. The blind spot is therefore published, not hypothesised. | fact | documented | §2 | "The vendor's own source paper reports that the ensemble property D(X) moves (inter-response diversity falls) while every measured Q(xᵢ) stays flat." | yes |
| C1555-03 | The joint-entropy contraction under a fixed key is the paper's own inference, kept separate from the published Self-BLEU fact. | interpretation | hypothesis | §2 | "The paper measures Self-BLEU, not joint entropy; the entropy-contraction reading is the information-theoretic interpretation this paper proposes, not the vendor's measurement." | yes (stated limit) |
| C1555-04 | Shumailov's collapse is tail-first and center-flat, which is exactly the profile per-instance benchmarks cannot see. | interpretation | attributed (Shumailov et al. 2024) + interpretation | §3, Leg 4 | "The center-flat, tail-dead signature is exactly the profile that per-instance benchmarks are structurally unable to see" | yes |
| C1555-05 | Sweeping a corpus for the signature cannot remove the model-level residue that fine-tuning transfers into weights. | function | interpretation | §5 | "later-generation corpora, the keyholder's own included, are downstream of models carrying residue no sweep of text can remove." | no |
| C1555-06 | Differential collapse exposure: if the keyholder uses privileged detection to clean its own corpora, the early generations' costs concentrate on the commons. | prediction | hypothesis ("contingent on practice") | §5 | "differential collapse exposure at the legible layer — if privileged, high-volume detection is used to clean the issuer's corpora while third-party access remains materially weaker, the early generations' auditable costs concentrate on the commons" | yes |
| C1555-07 | Tail-mass retention is a test with both outcomes stated: a positive result links the watermark to Shumailov's failure mode, and a null result separates the two. | falsification | hypothesis | §6, observable 2 | "A positive result would connect the watermark perturbation to the failure mode identified by Shumailov et al.; a null result would separate the phenomena" | yes |
| C1555-08 | The paper predicts ΔQ ≈ 0 and ΔD < 0 for marked versus unmarked generators of the same implementation. | prediction | hypothesis | §6, observable 1 | "Prediction from §2: ΔQ ≈ 0, ΔD < 0, small but signed." | no |
| C1555-09 | Limit: the paper does not claim that the watermark has caused measurable collapse. The watermark term is second-order next to model homogenisation. | falsification (stated limit) | self-description | §7; Operator's Review item 4 | "The Observatory does not claim the second compressor has caused measurable collapse." | yes |

REDUNDANCY: not fully redundant. #1556 is the declared companion ("companion to 'The Keyed Ensemble'") and restates C1555-01 (per-sequence certification against the ensemble, folded into its mismatched observation regime), C1555-05 (l.158: "corpus sweeping cannot remove the model-level residue"), C1555-07 (l.285, with the same positive/null wording) and the heritable-watermark leg. #1573 covers the benchmark-blindness side of C1555-04. Three things are not found in #1556 by search:
- C1555-06, differential collapse exposure on the commons;
- C1555-03, the Self-BLEU versus joint-entropy two-strength limit;
- C1555-09, the explicit disclaimer of demonstrated collapse (#1556 instead makes the watermark term "first-order in a contracted one").

#1555 is the earlier (origin) deposit of the shared claims.

---

## #1569 AXN-0661 — NOT
Reason: the Plato–Aristotle authorship notebook. "model collapse" appears 0 times, and there is no Shumailov and no recursive training. Every "collapse" is a measure failing or the Theaetetus wax block.
Best passage (l.2292): "Any C value for Plato–Aristotle establishes **relation**, never **authorship** — and the whole day is a record of measures that collapse at exactly that step."

## #1570 AXN-0662 — NOT
Reason: an authorship measurement register. "model collapse" appears 0 times, and "collapse" means a measure failing under controls.
Best passage (l.429): "But the matched 2×2 design collapses it: Plato(2)+Aristotle(2) 0.1151 sits *between* two independent controls"

## #1571 AXN-0663 — NOT
Reason: an authorship pre-registration ledger. "model collapse" appears 0 times.
Best passage (l.1163): "the whole day's work is the record of how many measures collapse at exactly that step."

## #1572 AXN-0664 — NOT
Reason: an authorship measures report. "model collapse" appears 0 times.
Best passage (l.436): "**Coverage geometry looked excellent and collapsed under contiguous size-matching**"

---

## #1611 AXN-069B — ADMIT
"The Negative of the Negative" (EA-NEGONT-01 v0.13). Model collapse is one of its two worked concepts. It states the standing representation (R_0) against the archive's (R_H), then makes its own claims:
- the exclusion of R_H is self-blinding and reflexive;
- composition around the lineage enacts collapse on collapse's own representation (negative bleed);
- the exclusion advances the phenomenon;
- the cost is denominated in capital, and irreversibility makes it unbounded;
- the model-collapse row carries a named test and states its limits.

| claim_id | claim | kind | modality | locus | verbatim | constitutive |
|---|---|---|---|---|---|---|
| C1611-01 | The standing representation of model collapse is single-substrate and single-recursion, measured by benchmarks. | interpretation | attributed (Shumailov et al. 2024) + interpretation | §1f | "The standing representation is single-substrate and single-recursion: a generative model trained on its own outputs loses the tails of its distribution (Shumailov et al. 2024), a pathology of artificial training loops, measured by benchmarks." | yes |
| C1611-02 | The archive's representation is one dynamical regime, regeneration against pruning with a bistable trap, operating in every substrate that processes language. | definition | attributed (to #855, #854, #1556, #1573, #1540, #191, #163, #1574) | §1f | "one dynamical regime — regeneration against pruning, with a bistable trap below a diversity threshold — operating in every substrate that processes language" | yes |
| C1611-03 | The 20/20 machine referee reports are disciplinary model collapse observed in the composition layer, which composes it as "systematic". | interpretation | interpretation | §1f | "twenty machine referee reports supplying an adjective the sentence does not contain, converging on the reception's mode regardless of the object — disciplinary model collapse observed in the composition layer" | no |
| C1611-04 | The exclusion is self-blinding. A graph that excludes R_H loses the reading of its own state variable, so it crosses the trap without a reading. | function | model consequence (hypothesis) | §1f | "A graph that excludes R_H at *model collapse* does not lose a paper; it loses the reading of its own state variable" | yes |
| C1611-05 | Negative bleed at model collapse is reflexive: composition around a tail source converges on the mode and prunes the mechanism, which is collapse enacted on its own representation. | function | hypothesis ("if the archive's representation is the correct one") | §1g | "Composition around a tail source converges on the mode and prunes the mechanism — R_H's own description of collapse, enacted on R_H at c." | yes |
| C1611-06 | The exclusion advances the phenomenon: it distorts the graph's instrument for detecting its own loss in the direction R_H predicts. | prediction | model consequence | §4f | "The exclusion is not neutral on the phenomenon it excludes the representation of. It advances it." | yes |
| C1611-07 | The cost of the exclusion is irreversibility, because below the threshold regeneration cannot outpace pruning. The exclusion is itself a tail-pruning event. | interpretation | model consequence | §4e | "the exclusion itself is a tail-pruning event of the kind R_H would have counted. In the phenomenon's terms the price has a form the earlier sections could not give it: irreversibility." | yes |
| C1611-08 | The cost is denominated in exposed capital: C = V·s·ΔP·L·u·η. ΔP and L are unmeasured, and the archive's instrument is uncalibrated. | function | hypothesis (assumed for dollar harm) | §5.4 | "Neither is measured; the archive's own instrument is a concept draft, uncalibrated by its own declaration, and the calibration run (produce collapse deliberately on a small model, watch the index) is among the things §4d says the scope keeps from being motivated." | yes (stated limit) |
| C1611-09 | Limit: the components are the published record and the synthesis is the claim. The instrument is provisional, and the row's status is "components: measured (external); synthesis: supported; instrument: provisional". | attribution | self-description | §7.1 table | "the synthesis is the claim and the instrument is uncalibrated by declaration" | yes |
| C1611-10 | Falsifier: the model-collapse row has no seated instance of negative bleed. The first battery is the four dimension-queries, predicting Q_s non-zero wherever the apparatus appears unsourced. | falsification | hypothesis | §8 | "the model-collapse row has none. The first untied battery on that row should be the four dimension-queries named in §1g, scored on Q against their retrieved basins, with the prediction that Q_s is non-zero wherever the substrate-agnostic or observation-regime apparatus appears unsourced." | yes |

REDUNDANCY: partial only. C1611-02 is redundant by attribution with #855 (substrates, frictionless path), #1556 (coupled recursions, observation regime), #1573 (wrong unit), #1 (classifier path), #1574 (disciplinary path). C1611-03 restates #1574. C1611-01 is a reading of Shumailov. C1611-04 through C1611-10 are not made by any listed deposit:
- reflexive negative bleed;
- the exclusion advancing collapse;
- irreversibility as the price;
- the capital-denominated cost model;
- the row's test and its limits.

#1616 (flattening that compounds by write-back) is the nearest neighbour to C1611-05 and C1611-06, but it is later and about write-back, not about the representation's exclusion. That is unverified here: #1616 was not opened.

---

## #1613 AXN-069F — ADMIT (borderline)
"What Not Reading Did to Its Own Ontology" (EA-MIRROR-01 v0.5). The phrase "model collapse" appears only in the keywords. The only engagement is §6, which extends #1573's head-sampling regime to a new substrate: a source-admission ontology's signals about its own health. It adds a prediction and states that the prediction can fail. Admit on that extension. If the ledger requires a claim about model collapse proper rather than an application of its instrument structure, this is the candidate to drop.

| claim_id | claim | kind | modality | locus | verbatim | constitutive |
|---|---|---|---|---|---|---|
| C1613-01 | Premise taken from #1573: a collapsing system's state variable declines in the tail while its instruments sample the head. | attribution | attributed (#1573) | §6 | "Under #1573 a collapsing system's state variable declines in the tail while its instruments sample the head." | no |
| C1613-02 | A source-admission ontology's diagnostics of its own health are head-sampling by construction, so they can fail to perceive tail loss. | function | model consequence (hypothesis, "in the modelled regime") | §6 | "a diagnostic that samples the admitted component is head-sampling by construction: it can fail to perceive tail loss." | yes |
| C1613-03 | Its damage therefore has no representation inside the ontology, because any text describing that damage enters under the source-relative barrier. | function | model consequence | §6 | "have no representation inside the ontology, because each would have to enter as a text describing the ontology, and such a text enters under §5." | yes |
| C1613-04 | Prediction and falsifier: the negative-of-the-negative dataset will be typed by its nearest existing type and not read. The test can fail, and the gate-reception test would falsify the barrier for an address. | prediction / falsification | hypothesis | §6; §7 | "the model's prediction is that it will be typed by its nearest existing type and not read. That prediction is this section stated as a test, and it can fail." | yes |
| C1613-05 | Limit: §§5–6 are model consequences until the two named tests exist. | falsification (stated limit) | self-description | §7 | "Until they exist, §§5–6 are model consequences and §§1 and 4 are scoped as marked." | yes |

REDUNDANCY: C1613-01 is redundant with #1573, the same claim attributed. C1613-02 through C1613-05 extend that claim to a new substrate (the admission ontology's self-diagnostics) and are not made by any listed deposit. #1611 C1611-04 (the graph loses the reading of its own state variable) is the nearest neighbour: same predicate, different subject (concept graph under exclusion versus admission ontology), and #1611 is not on the comparison list.
