| claim_id | claim | kind | modality | locus | verbatim quote | constitutive |
|---|---|---|---|---|---|---|
| L855-01 | Model collapse is a property of language, not of language models. | definition | hypothesis | Abstract | "This paper argues it is not a property of language models. It is a property of language." | yes |
| L855-02 | One dynamical regime operates in different substrates. | interpretation | hypothesis | Abstract | "This is not an analogy. It is the same dynamical regime operating in different substrates." | yes |
| L855-03 | Three substrates: silicon, adult human cognition, developing human cognition. | fact | attributed | Abstract | "In silicon: recursive training on model outputs produces measurable tail loss and distribution narrowing (Shumailov et al. 2024). In adult human cognition: sustained interaction with AI systems produces measurable homogenization of writing" | yes |
| L855-04 | The analogy is dynamical, not moral. | self-description | self-description | Abstract | "The analogy across substrates is dynamical, not moral: the paper does not claim that AI use is equivalent to child deprivation." | yes |
| L855-05 | The law is the same; reversibility is substrate-specific. | interpretation | hypothesis | §II The Structural Identity | "The dynamical law is the same; the reversibility parameter is substrate-specific." | yes |
| L855-06 | No current factor adds friction on the path to collapse. | prediction | hypothesis | Abstract | "there is not a single factor in the current development landscape that adds friction on the path to model collapse." | yes |
| L855-07 | The loss registers as improvement. | interpretation | interpretation | §II Substrate 2 | "The diversity contraction registers as improvement." | no |
| L855-08 | Feral-child evidence is historical and contested. | self-description | self-description | §II Substrate 3 | "the feral child cases are historical, not experimental, and the two most documented cases are contested." | yes |
| L855-09 | Trace's semiotic death drive is claimed as a prior formulation. | genealogy | interpretation | §V | "This is model collapse in semiotic vocabulary, stated in February 2026." | no |
| L855-10 | Falsifiers per substrate. | falsification | self-description | §VII | "The substrate-agnostic claim fails if any substrate demonstrates sustained capacity maintenance under input-diversity contraction without explicit diversity-injection design." | yes |
| L856-01 | Chat data is not a clean training source. | definition | hypothesis | Abstract | "Chat data fails the pristine test on three independent grounds." | yes |
| L856-02 | Training on chat data introduces collapse through an unmonitored channel. | prediction | hypothesis | Abstract | "It introduces model collapse through a channel no current benchmark, audit, or data-quality framework is measuring." | yes |
| L856-03 | The pristine source does not exist. | interpretation | hypothesis | Abstract | "The pristine source does not exist." | yes |
| L856-04 | Falsifier F1: no tail-thinning in heavy users' unaided writing. | falsification | self-description | §VI | "F1. If a well-powered study (N > 500, pre-registered) demonstrates that heavy AI-chat users show no statistically significant tail-thinning in their unaided writing relative to matched non-users, Ground 1 fails." | yes |
| L856-05 | Falsifier F2: no within-conversation convergence. | falsification | self-description | §VI | "F2. If turn-level analysis of a large chat corpus (> 100,000 conversations) shows no within-conversation convergence of user inputs toward model outputs on lexical, syntactic, or perplexity measures, Ground 2 fails." | yes |
| L856-06 | The studies have not been conducted. | self-description | self-description | §VI | "None of these studies has been conducted." | yes |
| L1573-01 | Turn-based benchmarks sample the head of the distribution. | interpretation | hypothesis | §I.2 | "each item is answered from the *head* of the model's distribution" | yes |
| L1573-02 | The collapse diagnostic should gate training. | function | hypothesis | §I.3 | "for the decision *what to train on next* the instrument reading the state variable should gate, and the instruments reading the head should inform." | yes |
| L1573-03 | Precision can rise while recall falls. | fact | attributed | §I.1 | "Alemohammad et al. (ICLR 2024): quality holds or improves while coverage falls." | no |
| L1573-04 | The instrument is uncalibrated and untested. | self-description | self-description | front matter | "The instrument is specified and NOT calibrated, NOT tested, NOT run" | yes |
| L855-11 | The three substrates differ in severity, mechanism and timescale but are governed by the same dynamical law. | interpretation | hypothesis | §II The Structural Identity | "The three cases differ in severity, in mechanism, and in timescale. But they are governed by the same dynamical law" | yes |
