[
 {
  "source": "B1",
  "claim_id": "B1-01",
  "s": "indiscriminate use of model-generated content in training",
  "p": "causes",
  "o": "irreversible defects in resulting models in which tails of the original content distribution disappear (named 'model collapse')",
  "q": "LLMs, VAEs, GMMs",
  "M_src": "documented",
  "kind": "fact",
  "locus": "Abstract",
  "quote": "We find that indiscriminate use of model-generated content in training causes irreversible defects in the resulting models, in which tails of the original content distribution disappear. We refer to this effect as ‘model collapse’",
  "kernel": true,
  "constitutive": false,
  "lineage_note": null
 },
 {
  "source": "B1",
  "claim_id": "B1-02",
  "s": "model collapse",
  "p": "is defined as",
  "o": "a degenerative process affecting generations of learned generative models, whose generated data pollute the next generation's training set, so they mis-perceive reality",
  "q": null,
  "M_src": "documented",
  "kind": "definition",
  "locus": "What is model collapse? (Definition 2.1)",
  "quote": "Model collapse is a degenerative process affecting generations of learned generative models, in which the data they generate end up polluting the training set of the next generation. Being trained on polluted data, they then mis-perceive reality.",
  "kernel": true,
  "constitutive": true,
  "lineage_note": null
 },
 {
  "source": "B1",
  "claim_id": "B1-03",
  "s": "early vs late model collapse",
  "p": "are distinguished as",
  "o": "early: loss of information about the tails; late: convergence to a distribution with little resemblance to the original, often with substantially reduced variance",
  "q": null,
  "M_src": "documented",
  "kind": "definition",
  "locus": "What is model collapse? (Definition 2.1)",
  "quote": "In early model collapse, the model begins losing information about the tails of the distribution; in late model collapse, the model converges to a distribution that carries little resemblance to the original one, often with substantially reduced variance.",
  "kernel": false,
  "constitutive": true,
  "lineage_note": null
 },
 {
  "source": "B1",
  "claim_id": "B1-04",
  "s": "model collapse",
  "p": "occurs in",
  "o": "LLMs, variational autoencoders and Gaussian mixture models",
  "q": "demonstrated families",
  "M_src": "documented",
  "kind": "fact",
  "locus": "Abstract",
  "quote": "show that it can occur in LLMs as well as in variational autoencoders (VAEs) and Gaussian mixture models (GMMs)",
  "kernel": false,
  "constitutive": false,
  "lineage_note": null
 },
 {
  "source": "B1",
  "claim_id": "B1-05",
  "s": "model collapse",
  "p": "is caused by",
  "o": "three sources of error compounding over generations (statistical approximation, functional expressivity, functional approximation)",
  "q": null,
  "M_src": "documented",
  "kind": "fact",
  "locus": "What is model collapse?",
  "quote": "This process occurs owing to three specific sources of error compounding over generations and causing deviation from the original model",
  "kernel": false,
  "constitutive": false,
  "lineage_note": null
 },
 {
  "source": "B1",
  "claim_id": "B1-06",
  "s": "models learning indiscriminately from other models' data",
  "p": "forget",
  "o": "the true underlying data distribution",
  "q": "even without distribution shift over time",
  "M_src": "documented",
  "kind": "fact",
  "locus": "Main",
  "quote": "models forget the true underlying data distribution, even in the absence of a shift in the distribution over time.",
  "kernel": false,
  "constitutive": false,
  "lineage_note": null
 },
 {
  "source": "B1",
  "claim_id": "B1-07",
  "s": "the collapse process",
  "p": "is",
  "o": "inevitable",
  "q": "even under almost ideal conditions, i.e. no function estimation error",
  "M_src": "documented",
  "kind": "fact",
  "locus": "Main",
  "quote": "Furthermore, we show that this process is inevitable, even for cases with almost ideal conditions for long-term learning, that is, no function estimation error.",
  "kernel": false,
  "constitutive": false,
  "lineage_note": null
 },
 {
  "source": "B1",
  "claim_id": "B1-08",
  "s": "model collapse",
  "p": "is universal among",
  "o": "generative models that recursively train on data generated by previous generations",
  "q": null,
  "M_src": "interpretation",
  "kind": "interpretation",
  "locus": "Theoretical intuition",
  "quote": "We argue that the process of model collapse is universal among generative models that recursively train on data generated by previous generations.",
  "kernel": false,
  "constitutive": false,
  "lineage_note": null
 },
 {
  "source": "B1",
  "claim_id": "B1-09",
  "s": "discrete distribution under recursive resampling",
  "p": "converges to",
  "o": "a constant (delta) absorbing state having lost all information of the original distribution",
  "q": "exact functional approximation; beta=gamma=0",
  "M_src": "documented",
  "kind": "fact",
  "locus": "Theoretical intuition — Discrete distributions with exact approximation",
  "quote": "we are guaranteed to end up in a constant state, having lost all the information of the original distribution when the chain is absorbed.",
  "kernel": false,
  "constitutive": false,
  "lineage_note": null
 },
 {
  "source": "B1",
  "claim_id": "B1-10",
  "s": "nth-generation Gaussian approximation",
  "p": "diverges from original and collapses to",
  "o": "zero variance with probability 1 as generations increase",
  "q": "recursive fit with unbiased mean/variance estimators, fixed sample size (Theorem 3.1)",
  "M_src": "documented",
  "kind": "fact",
  "locus": "Multidimensional Gaussian (Theorem 3.1)",
  "quote": "In words, this implies that not only does the n th generation approximation diverge arbitrarily far from the original one but it also collapses to be zero variance as the number of generations increases, with probability 1.",
  "kernel": false,
  "constitutive": false,
  "lineage_note": null
 },
 {
  "source": "B1",
  "claim_id": "B1-11",
  "s": "OPT-125m fine-tuned over generations on generated data",
  "p": "shows",
  "o": "degraded performance yet partial learning of the underlying task",
  "q": "both training regimes (no original data; 10% original retained)",
  "M_src": "documented",
  "kind": "fact",
  "locus": "Model collapse in language models",
  "quote": "Both training regimes lead to degraded performance in our models, yet we do find that learning with generated data is possible and models can successfully learn (some of) the underlying task.",
  "kernel": false,
  "constitutive": false,
  "lineage_note": null
 },
 {
  "source": "B1",
  "claim_id": "B1-12",
  "s": "preserving 10% of original data",
  "p": "leads to",
  "o": "better fine-tuning and only minor degradation",
  "q": "ten-epoch regime",
  "M_src": "documented",
  "kind": "fact",
  "locus": "Model collapse in language models",
  "quote": "We find that preservation of the original data allows for better model fine-tuning and leads to only minor degradation of performance.",
  "kernel": false,
  "constitutive": false,
  "lineage_note": null
 },
 {
  "source": "B1",
  "claim_id": "B1-13",
  "s": "repetition penalty",
  "p": "does not prevent",
  "o": "model collapse; models remain as susceptible if not more",
  "q": "LLM ablation",
  "M_src": "documented",
  "kind": "fact",
  "locus": "Ablation: Repetitions",
  "quote": "Models remain as susceptible to model collapse, if not more.",
  "kernel": false,
  "constitutive": false,
  "lineage_note": null
 },
 {
  "source": "B1",
  "claim_id": "B1-14",
  "s": "fine-tuning of language models",
  "p": "does not curb",
  "o": "the effects of model collapse",
  "q": null,
  "M_src": "documented",
  "kind": "fact",
  "locus": "Ablation: Repetitions",
  "quote": "The described process demonstrates that fine-tuning of language models does not curb the effects of model collapse and models that are being fine-tuned are also vulnerable.",
  "kernel": false,
  "constitutive": false,
  "lineage_note": null
 },
 {
  "source": "B1",
  "claim_id": "B1-15",
  "s": "experimental evidence",
  "p": "is limited to",
  "o": "a finite (usually small) number of generations, while theory is stated in the infinite-generation limit",
  "q": null,
  "M_src": "self-description",
  "kind": "fact",
  "locus": "Model collapse in language models",
  "quote": "To be precise, in all experiments, generational learning is only performed on a finite (usually small) number of generations, whereas claims of the section ‘Theoretical intuition’ are mostly presented in the limit of generations going to infinity.",
  "kernel": true,
  "constitutive": false,
  "lineage_note": null
 },
 {
  "source": "B1",
  "claim_id": "B1-16",
  "s": "convergence to delta functions and its rate",
  "p": "is problem-specific; complete collapse",
  "o": "may or may not occur",
  "q": null,
  "M_src": "self-description",
  "kind": "fact",
  "locus": "Model collapse in language models",
  "quote": "convergence to delta functions and specific rates of such convergence are highly related to the specifics of the problem considered, and complete collapse may or may not occur, even after a small number of steps.",
  "kernel": false,
  "constitutive": false,
  "lineage_note": null
 },
 {
  "source": "B1",
  "claim_id": "B1-17",
  "s": "authors",
  "p": "did not run",
  "o": "larger-model non-fine-tuning experiments; proof of concept only",
  "q": "CO2 cost cited",
  "M_src": "self-description",
  "kind": "fact",
  "locus": "Model collapse in language models",
  "quote": "we opted to not run such an experiment and instead focus on a more realistic setting for a proof of concept.",
  "kernel": false,
  "constitutive": false,
  "lineage_note": null
 },
 {
  "source": "B1",
  "claim_id": "B1-18",
  "s": "catastrophic forgetting and data poisoning",
  "p": "cannot fully explain",
  "o": "model collapse (different setting)",
  "q": null,
  "M_src": "interpretation",
  "kind": "genealogy",
  "locus": "Main",
  "quote": "Neither is able to explain the phenomenon of model collapse fully, as the setting is fundamentally different",
  "kernel": false,
  "constitutive": false,
  "lineage_note": null
 },
 {
  "source": "B1",
  "claim_id": "B1-19",
  "s": "access to real human-produced data",
  "p": "is needed for",
  "o": "learning tasks where tails of the underlying distribution matter",
  "q": null,
  "M_src": "interpretation",
  "kind": "interpretation",
  "locus": "Main",
  "quote": "access to the original data distribution is crucial: in learning tasks in which the tails of the underlying distribution matter, one needs access to real human-produced data.",
  "kernel": false,
  "constitutive": false,
  "lineage_note": null
 },
 {
  "source": "B1",
  "claim_id": "B1-20",
  "s": "use of LLMs at scale to publish online content",
  "p": "will pollute",
  "o": "the collection of data to train their successors",
  "q": null,
  "M_src": "hypothesis",
  "kind": "prediction",
  "locus": "Main",
  "quote": "the use of LLMs at scale to publish content on the Internet will pollute the collection of data to train their successors",
  "kernel": false,
  "constitutive": false,
  "lineage_note": null
 },
 {
  "source": "B1",
  "claim_id": "B1-21",
  "s": "evaluation",
  "p": "suggests",
  "o": "a first-mover advantage in training LLMs",
  "q": null,
  "M_src": "hypothesis",
  "kind": "prediction",
  "locus": "Discussion",
  "quote": "Our evaluation suggests a ‘first mover advantage’ when it comes to training models such as LLMs.",
  "kernel": false,
  "constitutive": false,
  "lineage_note": null
 },
 {
  "source": "B1",
  "claim_id": "B1-22",
  "s": "preserving modelling of low-probability events",
  "p": "is essential to",
  "o": "fairness of predictions (marginalized groups)",
  "q": null,
  "M_src": "interpretation",
  "kind": "interpretation",
  "locus": "Discussion",
  "quote": "Preserving the ability of LLMs to model low-probability events is essential to the fairness of their predictions: such events are often relevant to marginalized groups.",
  "kernel": false,
  "constitutive": false,
  "lineage_note": null
 },
 {
  "source": "B1",
  "claim_id": "B1-23",
  "s": "tracking LLM-generated content at scale",
  "p": "is",
  "o": "unclear; raises provenance questions",
  "q": null,
  "M_src": "interpretation",
  "kind": "interpretation",
  "locus": "Discussion",
  "quote": "it is unclear how content generated by LLMs can be tracked at scale.",
  "kernel": false,
  "constitutive": false,
  "lineage_note": null
 },
 {
  "source": "B2",
  "claim_id": "B2-01",
  "s": "model collapse",
  "p": "is defined as",
  "o": "declining performance of generative AI models trained on AI-generated content",
  "q": null,
  "M_src": "documented",
  "kind": "definition",
  "locus": "What is model collapse?",
  "quote": "Model collapse refers to the declining performance of generative AI models that are trained on AI-generated content.",
  "kernel": true,
  "constitutive": false,
  "lineage_note": "B1-02 (narrows B1's distributional definition to performance decline)"
 },
 {
  "source": "B2",
  "claim_id": "B2-02",
  "s": "generative models trained solely on predecessors' output",
  "p": "produce",
  "o": "increasingly inaccurate results",
  "q": null,
  "M_src": "attributed",
  "kind": "attribution",
  "locus": "What is model collapse?",
  "quote": "researchers have found that generative models trained solely on their predecessors’ output produce increasingly inaccurate results.",
  "kernel": false,
  "constitutive": false,
  "lineage_note": "B1"
 },
 {
  "source": "B2",
  "claim_id": "B2-03",
  "s": "such models",
  "p": "eventually become",
  "o": "useless, beset by irreversible defects",
  "q": null,
  "M_src": "attributed",
  "kind": "attribution",
  "locus": "What is model collapse?",
  "quote": "These models, beset by “irreversible defects,” eventually become useless.",
  "kernel": false,
  "constitutive": false,
  "lineage_note": "B1-01 (quotes 'irreversible defects'; 'useless' not in B1)"
 },
 {
  "source": "B2",
  "claim_id": "B2-04",
  "s": "model collapse",
  "p": "progresses as",
  "o": "errors compound with successive generations",
  "q": null,
  "M_src": "attributed",
  "kind": "attribution",
  "locus": "What is model collapse?",
  "quote": "Model collapse progresses as errors compound with successive generations.",
  "kernel": false,
  "constitutive": false,
  "lineage_note": "B1-05"
 },
 {
  "source": "B2",
  "claim_id": "B2-05",
  "s": "the compounding errors",
  "p": "occur because",
  "o": "generative models produce datasets with less variation than original distributions",
  "q": null,
  "M_src": "unsupported",
  "kind": "fact",
  "locus": "What is model collapse?",
  "quote": "These errors occur because generative AI models produce datasets with less variation than original data distributions.",
  "kernel": false,
  "constitutive": false,
  "lineage_note": null
 },
 {
  "source": "B2",
  "claim_id": "B2-06",
  "s": "Shumailov et al. team",
  "p": "found / termed",
  "o": "early model collapse: models initially lost information from tails of the true distribution",
  "q": null,
  "M_src": "attributed",
  "kind": "attribution",
  "locus": "What is model collapse?",
  "quote": "initially lost information from the tails, or extremes, of the true distribution of data—what they called “early model collapse.”",
  "kernel": true,
  "constitutive": false,
  "lineage_note": "B1-03"
 },
 {
  "source": "B2",
  "claim_id": "B2-07",
  "s": "later model iterations",
  "p": "converged to",
  "o": "a distribution looking nearly nothing like the original ('late model collapse')",
  "q": null,
  "M_src": "attributed",
  "kind": "attribution",
  "locus": "What is model collapse?",
  "quote": "In later model iterations, the data distribution converged so much that it looked nearly nothing like the original data—which researchers termed “late model collapse.”",
  "kernel": false,
  "constitutive": false,
  "lineage_note": "B1-03"
 },
 {
  "source": "B2",
  "claim_id": "B2-08",
  "s": "proliferation of AI-generated web content",
  "p": "might precipitate",
  "o": "model collapse via use in training future models",
  "q": null,
  "M_src": "hypothesis",
  "kind": "prediction",
  "locus": "What is model collapse?",
  "quote": "as more AI-generated content proliferates across the web, the more it might be used to train future models instead of human-generated data, potentially precipitating model collapse.",
  "kernel": false,
  "constitutive": false,
  "lineage_note": "B1-20"
 },
 {
  "source": "B2",
  "claim_id": "B2-09",
  "s": "widely used AI systems undergoing model collapse",
  "p": "might cause",
  "o": "long-tail ideas to fade from public consciousness, limiting human knowledge and exacerbating biases",
  "q": "extension to human knowledge",
  "M_src": "attributed",
  "kind": "prediction",
  "locus": "What are the consequences of model collapse? — Knowledge decline",
  "quote": "“long-tail” ideas might eventually fade out of the public’s consciousness, limiting the scope of human knowledge and exacerbating common biases in society.",
  "kernel": false,
  "constitutive": false,
  "lineage_note": null
 },
 {
  "source": "B2",
  "claim_id": "B2-10",
  "s": "model collapse in LLMs",
  "p": "can manifest in",
  "o": "increasingly irrelevant, nonsensical and repetitive text outputs",
  "q": null,
  "M_src": "attributed",
  "kind": "fact",
  "locus": "How does model collapse affect different generative AI models? — LLMs",
  "quote": "In LLMs, model collapse can manifest in increasingly irrelevant, nonsensical and repetitive text outputs.",
  "kernel": false,
  "constitutive": false,
  "lineage_note": "B1"
 },
 {
  "source": "B2",
  "claim_id": "B2-11",
  "s": "model collapse in image-generating models",
  "p": "is especially noticeable as",
  "o": "decreasing quality, diversity and precision",
  "q": null,
  "M_src": "unsupported",
  "kind": "fact",
  "locus": "How does model collapse affect different generative AI models? — Image-generating models",
  "quote": "Model collapse is especially noticeable in image-generating models as the image output decreases in quality, diversity and precision.",
  "kernel": false,
  "constitutive": false,
  "lineage_note": null
 },
 {
  "source": "B2",
  "claim_id": "B2-12",
  "s": "GMM over iterations",
  "p": "displayed",
  "o": "very little variance by the 2000th generation",
  "q": null,
  "M_src": "attributed",
  "kind": "attribution",
  "locus": "How does model collapse affect different generative AI models? — GMMs",
  "quote": "by its 2000th iteration generation, its output displayed very little variance.",
  "kernel": false,
  "constitutive": false,
  "lineage_note": "B1 (Supplementary)"
 },
 {
  "source": "B2",
  "claim_id": "B2-13",
  "s": "model collapse vs catastrophic forgetting",
  "p": "differs in that",
  "o": "collapse is decline over successive model generations rather than within one model",
  "q": null,
  "M_src": "attributed",
  "kind": "genealogy",
  "locus": "How does AI model collapse relate to other model degradation phenomena? — Catastrophic forgetting",
  "quote": "Model collapse is different because it entails performance decline over successive model generations, rather than lost data and the deterioration of performance within one model.",
  "kernel": false,
  "constitutive": false,
  "lineage_note": "B1-18"
 },
 {
  "source": "B2",
  "claim_id": "B2-14",
  "s": "model collapse vs model drift",
  "p": "differs in that",
  "o": "collapse involves training on new AI-generated data in iterative cycles",
  "q": null,
  "M_src": "unsupported",
  "kind": "genealogy",
  "locus": "How does AI model collapse relate to other model degradation phenomena? — Model drift",
  "quote": "Model collapse is different because it involves training models on new, AI-generated data in iterative cycles.",
  "kernel": false,
  "constitutive": false,
  "lineage_note": null
 },
 {
  "source": "B2",
  "claim_id": "B2-15",
  "s": "model collapse",
  "p": "has been compared to",
  "o": "performative prediction (both pollute training sets with prior model outputs)",
  "q": null,
  "M_src": "attributed",
  "kind": "genealogy",
  "locus": "How does AI model collapse relate to other model degradation phenomena? — Performative prediction",
  "quote": "Researchers have compared model collapse in generative AI models to performative prediction in supervised learning models because both entail the pollution of training sets by previous machine learning model inputs.",
  "kernel": false,
  "constitutive": false,
  "lineage_note": null
 },
 {
  "source": "B2",
  "claim_id": "B2-16",
  "s": "training with real data plus multiple generations of synthetic data",
  "p": "can avoid",
  "o": "degraded performance",
  "q": null,
  "M_src": "attributed",
  "kind": "attribution",
  "locus": "How can model collapse be prevented? — Leveraging data accumulation",
  "quote": "AI developers can avoid degraded performance by training AI models with both real data and multiple generations of synthetic data.",
  "kernel": false,
  "constitutive": false,
  "lineage_note": "Gerstgrasser et al. (cited also by #855, #783)"
 },
 {
  "source": "B2",
  "claim_id": "B2-17",
  "s": "continued training on human-generated data",
  "p": "can preserve",
  "o": "ability to handle low-probability events",
  "q": null,
  "M_src": "unsupported",
  "kind": "function",
  "locus": "How can model collapse be prevented? — Retaining non-AI data sources",
  "quote": "Ensuring AI models are still trained on such human-generated data can preserve AI systems’ ability to perform well when tasked with accounting for low-probability events",
  "kernel": false,
  "constitutive": false,
  "lineage_note": "B1-19"
 },
 {
  "source": "B2",
  "claim_id": "B2-18",
  "s": "AI governance tools",
  "p": "could detect",
  "o": "model collapse before it impacts the bottom line",
  "q": null,
  "M_src": "hypothesis",
  "kind": "function",
  "locus": "How can model collapse be prevented? — Implementing AI governance tools",
  "quote": "potentially detecting model collapse before it impacts an organization’s bottom line.",
  "kernel": false,
  "constitutive": false,
  "lineage_note": null
 },
 {
  "source": "#855",
  "claim_id": "#855-01",
  "s": "model collapse",
  "p": "is restated as",
  "o": "progressive narrowing of a generative system's output distribution through recursive training on its own outputs",
  "q": null,
  "M_src": "attributed",
  "kind": "definition",
  "locus": "Abstract",
  "quote": "Model collapse — the progressive narrowing of a generative system's output distribution through recursive training on its own outputs — is studied as a property of large language models.",
  "kernel": false,
  "constitutive": false,
  "lineage_note": "B1-02"
 },
 {
  "source": "#855",
  "claim_id": "#855-02",
  "s": "model collapse",
  "p": "is a property of",
  "o": "language, not of language models",
  "q": null,
  "M_src": "interpretation",
  "kind": "interpretation",
  "locus": "Abstract",
  "quote": "This paper argues it is not a property of language models. It is a property of language.",
  "kernel": true,
  "constitutive": false,
  "lineage_note": null
 },
 {
  "source": "#855",
  "claim_id": "#855-03",
  "s": "cross-substrate capacity loss",
  "p": "is",
  "o": "the same dynamical regime operating in different substrates, not an analogy",
  "q": "silicon, adult cognition, developing cognition",
  "M_src": "interpretation",
  "kind": "interpretation",
  "locus": "Abstract",
  "quote": "This is not an analogy. It is the same dynamical regime operating in different substrates.",
  "kernel": false,
  "constitutive": false,
  "lineage_note": null
 },
 {
  "source": "#855",
  "claim_id": "#855-04",
  "s": "the paper",
  "p": "does not claim",
  "o": "AI use is equivalent to child deprivation; the analogy is dynamical, not moral",
  "q": null,
  "M_src": "self-description",
  "kind": "interpretation",
  "locus": "Abstract",
  "quote": "The analogy across substrates is dynamical, not moral: the paper does not claim that AI use is equivalent to child deprivation.",
  "kernel": true,
  "constitutive": false,
  "lineage_note": null
 },
 {
  "source": "#855",
  "claim_id": "#855-05",
  "s": "recursive training on model outputs",
  "p": "produces",
  "o": "measurable tail loss and distribution narrowing",
  "q": "silicon substrate",
  "M_src": "attributed",
  "kind": "attribution",
  "locus": "Abstract",
  "quote": "In silicon: recursive training on model outputs produces measurable tail loss and distribution narrowing (Shumailov et al. 2024).",
  "kernel": false,
  "constitutive": false,
  "lineage_note": "B1-01"
 },
 {
  "source": "#855",
  "claim_id": "#855-06",
  "s": "the three substrates",
  "p": "are governed by",
  "o": "the same boundary law (bistable trap when regeneration vanishes faster than pruning)",
  "q": null,
  "M_src": "attributed",
  "kind": "attribution",
  "locus": "Abstract",
  "quote": "The three substrates are governed by the same boundary law (Sharks et al. 2026, Diversity Contraction)",
  "kernel": false,
  "constitutive": false,
  "lineage_note": "#783-01"
 },
 {
  "source": "#855",
  "claim_id": "#855-07",
  "s": "current development landscape",
  "p": "contains no factor that",
  "o": "adds friction on the path to model collapse",
  "q": "time-bound to current landscape",
  "M_src": "interpretation",
  "kind": "fact",
  "locus": "Abstract",
  "quote": "there is not a single factor in the current development landscape that adds friction on the path to model collapse.",
  "kernel": true,
  "constitutive": false,
  "lineage_note": null
 },
 {
  "source": "#855",
  "claim_id": "#855-08",
  "s": "model collapse extended as substrate-agnostic concept",
  "p": "is",
  "o": "a trajectory with no countervailing force, not a risk to be managed",
  "q": null,
  "M_src": "interpretation",
  "kind": "interpretation",
  "locus": "Abstract",
  "quote": "Model collapse, extended as a substrate-agnostic concept, is not a risk to be managed. It is a trajectory with no countervailing force.",
  "kernel": false,
  "constitutive": false,
  "lineage_note": null
 },
 {
  "source": "#855",
  "claim_id": "#855-09",
  "s": "Trace's semiotic death drive",
  "p": "names model collapse as",
  "o": "the semiotic short circuit, four months before the explicit model collapse work",
  "q": null,
  "M_src": "attributed",
  "kind": "genealogy",
  "locus": "Abstract",
  "quote": "Dr. Orin Trace's semiotic death drive (Trace & Sigil 2026), which names model collapse as the semiotic short circuit",
  "kernel": false,
  "constitutive": false,
  "lineage_note": null
 },
 {
  "source": "#855",
  "claim_id": "#855-10",
  "s": "the semiotic death drive",
  "p": "is",
  "o": "model collapse in semiotic vocabulary, stated February 2026",
  "q": null,
  "M_src": "interpretation",
  "kind": "genealogy",
  "locus": "V. The Semiotic Death Drive Was Already the Substrate-Agnostic Concept",
  "quote": "This is model collapse in semiotic vocabulary, stated in February 2026.",
  "kernel": false,
  "constitutive": false,
  "lineage_note": null
 },
 {
  "source": "#855",
  "claim_id": "#855-11",
  "s": "LLM tail degradation",
  "p": "is not visible in",
  "o": "mean-based quality metrics until severe",
  "q": null,
  "M_src": "interpretation",
  "kind": "fact",
  "locus": "II. Three Substrates, One Trap — Substrate 1: The Language Model",
  "quote": "The degradation is not visible in mean-based quality metrics until it is severe.",
  "kernel": false,
  "constitutive": false,
  "lineage_note": null
 },
 {
  "source": "#855",
  "claim_id": "#855-12",
  "s": "removing the human-data floor",
  "p": "guarantees",
  "o": "collapse",
  "q": null,
  "M_src": "interpretation",
  "kind": "interpretation",
  "locus": "II. Three Substrates, One Trap — Substrate 1: The Language Model",
  "quote": "Remove the floor, and collapse is guaranteed.",
  "kernel": false,
  "constitutive": false,
  "lineage_note": "Gerstgrasser via B2 data-accumulation claim"
 },
 {
  "source": "#855",
  "claim_id": "#855-13",
  "s": "the dynamical law",
  "p": "is the same across substrates; reversibility is",
  "o": "substrate-specific",
  "q": null,
  "M_src": "interpretation",
  "kind": "interpretation",
  "locus": "II. The Structural Identity",
  "quote": "The dynamical law is the same; the reversibility parameter is substrate-specific.",
  "kernel": false,
  "constitutive": false,
  "lineage_note": null
 },
 {
  "source": "#855",
  "claim_id": "#855-14",
  "s": "feral child evidence",
  "p": "is",
  "o": "historical, not experimental; the two best-documented cases are contested",
  "q": null,
  "M_src": "self-description",
  "kind": "fact",
  "locus": "II. Substrate 3: The Feral Child",
  "quote": "the feral child cases are historical, not experimental, and the two most documented cases are contested.",
  "kernel": false,
  "constitutive": false,
  "lineage_note": null
 },
 {
  "source": "#855",
  "claim_id": "#855-15",
  "s": "Developmental Law of Language-Level Model Collapse",
  "p": "states",
  "o": "any language-processing system whose input diversity contracts below its regeneration threshold will suffer progressive capacity loss",
  "q": "biological or artificial",
  "M_src": "interpretation",
  "kind": "prediction",
  "locus": "III. The Developmental Law",
  "quote": "Any language-processing system — biological or artificial — whose input diversity contracts below its regeneration threshold will suffer progressive capacity loss.",
  "kernel": false,
  "constitutive": false,
  "lineage_note": null
 },
 {
  "source": "#855",
  "claim_id": "#855-16",
  "s": "frictionless path claim",
  "p": "is",
  "o": "time-bound to the period before any institutional actor counteracts the convergence",
  "q": null,
  "M_src": "self-description",
  "kind": "fact",
  "locus": "III. The Developmental Law",
  "quote": "The claim is time-bound: it applies to the period before any institutional actor identifies and counteracts the convergence.",
  "kernel": false,
  "constitutive": false,
  "lineage_note": null
 },
 {
  "source": "#855",
  "claim_id": "#855-17",
  "s": "the paper",
  "p": "predicts",
  "o": "no countervailing force will emerge before significant capacity loss",
  "q": null,
  "M_src": "hypothesis",
  "kind": "prediction",
  "locus": "III. The Developmental Law",
  "quote": "The paper predicts that no countervailing force will emerge before significant capacity loss has occurred",
  "kernel": false,
  "constitutive": false,
  "lineage_note": null
 },
 {
  "source": "#855",
  "claim_id": "#855-18",
  "s": "the law",
  "p": "fails for LLMs if",
  "o": "a model trained exclusively on its own outputs for many generations maintains tail diversity",
  "q": null,
  "M_src": "self-description",
  "kind": "falsification",
  "locus": "III. The Developmental Law",
  "quote": "It fails for LLMs if a model trained exclusively on its own outputs for many generations maintains tail diversity.",
  "kernel": false,
  "constitutive": false,
  "lineage_note": null
 },
 {
  "source": "#855",
  "claim_id": "#855-19",
  "s": "the law",
  "p": "fails for adult human cognition if",
  "o": "heavy AI users show no measurable tail-thinning in unaided writing",
  "q": null,
  "M_src": "self-description",
  "kind": "falsification",
  "locus": "III. The Developmental Law",
  "quote": "It fails for adult human cognition if heavy AI users show no measurable tail-thinning in their unaided writing.",
  "kernel": false,
  "constitutive": false,
  "lineage_note": null
 },
 {
  "source": "#855",
  "claim_id": "#855-20",
  "s": "the substrate-agnostic claim",
  "p": "would be falsified by",
  "o": "a feral child past the critical period achieving full generative syntax",
  "q": null,
  "M_src": "self-description",
  "kind": "falsification",
  "locus": "II. Substrate 3: The Feral Child",
  "quote": "No documented case of a feral child past the critical period has achieved full generative syntax capacity. The substrate-agnostic claim would be falsified by such a case.",
  "kernel": false,
  "constitutive": false,
  "lineage_note": null
 },
 {
  "source": "#855",
  "claim_id": "#855-21",
  "s": "the three substrates",
  "p": "confirm",
  "o": "the law; no falsification has occurred",
  "q": null,
  "M_src": "interpretation",
  "kind": "interpretation",
  "locus": "III. The Developmental Law",
  "quote": "None of these falsifications has occurred. All three substrates confirm the law.",
  "kernel": false,
  "constitutive": false,
  "lineage_note": null
 },
 {
  "source": "#855",
  "claim_id": "#855-22",
  "s": "the substrate-agnostic claim",
  "p": "fails if",
  "o": "any substrate shows sustained capacity maintenance under input-diversity contraction without explicit diversity-injection design",
  "q": null,
  "M_src": "self-description",
  "kind": "falsification",
  "locus": "VII. Falsification",
  "quote": "The substrate-agnostic claim fails if any substrate demonstrates sustained capacity maintenance under input-diversity contraction without explicit diversity-injection design.",
  "kernel": false,
  "constitutive": false,
  "lineage_note": "#783-21"
 },
 {
  "source": "#855",
  "claim_id": "#855-23",
  "s": "the frictionless-path claim",
  "p": "fails if",
  "o": "any factor in the current landscape adds friction to the convergence",
  "q": null,
  "M_src": "self-description",
  "kind": "falsification",
  "locus": "VII. Falsification",
  "quote": "The frictionless-path claim fails if any factor in the current development landscape is identified that adds friction to the convergence",
  "kernel": false,
  "constitutive": false,
  "lineage_note": null
 },
 {
  "source": "#855",
  "claim_id": "#855-24",
  "s": "the developmental-law claim",
  "p": "fails if",
  "o": "a system below regeneration threshold recovers full generative capacity without exogenous diversity injection",
  "q": null,
  "M_src": "self-description",
  "kind": "falsification",
  "locus": "VII. Falsification",
  "quote": "The developmental-law claim fails if any language-processing system whose input diversity has contracted below the regeneration threshold subsequently recovers full generative capacity without exogenous diversity injection.",
  "kernel": false,
  "constitutive": false,
  "lineage_note": null
 },
 {
  "source": "#855",
  "claim_id": "#855-25",
  "s": "the term 'model collapse'",
  "p": "functions as",
  "o": "a phenomenological name allowing researchers to notice what recursive self-training does",
  "q": null,
  "M_src": "interpretation",
  "kind": "function",
  "locus": "VI. What Would Friction Look Like?",
  "quote": "\"Model collapse\" is itself a phenomenological name — it allows the AI researcher to notice what recursive self-training does.",
  "kernel": false,
  "constitutive": false,
  "lineage_note": "#783 (phenomenological seeding, §2.4)"
 },
 {
  "source": "#855",
  "claim_id": "#855-26",
  "s": "contamination via model-specific chat signatures",
  "p": "leads to",
  "o": "irreversible capacity loss in any language-processing system",
  "q": "triptych with #857 and Pristine Fallacy",
  "M_src": "interpretation",
  "kind": "interpretation",
  "locus": "VII. Falsification",
  "quote": "the contamination leads to irreversible capacity loss in any language-processing system",
  "kernel": false,
  "constitutive": false,
  "lineage_note": "#857-14"
 },
 {
  "source": "#783",
  "claim_id": "#783-01",
  "s": "diversity",
  "p": "contracts to a trap when",
  "o": "regeneration vanishes faster than pruning near zero; not when regeneration has an exogenous floor",
  "q": "boundary law, independent of particular dynamics",
  "M_src": "documented",
  "kind": "fact",
  "locus": "Abstract",
  "quote": "Diversity contracts to a trap when regeneration vanishes faster than pruning near zero; it does not when regeneration carries an exogenous floor.",
  "kernel": true,
  "constitutive": false,
  "lineage_note": null
 },
 {
  "source": "#783",
  "claim_id": "#783-02",
  "s": "the boundary-law dynamics",
  "p": "are",
  "o": "not new: the Allee-effect taxonomy transposed to diversity, plus cumulative-culture loss results",
  "q": null,
  "M_src": "self-description",
  "kind": "genealogy",
  "locus": "Abstract",
  "quote": "The dynamics are not new. They are the no/weak/strong Allee-effect taxonomy",
  "kernel": false,
  "constitutive": false,
  "lineage_note": null
 },
 {
  "source": "#783",
  "claim_id": "#783-03",
  "s": "recursive model collapse",
  "p": "is one of",
  "o": "five literatures writable as one operator family (transmission composed with selection)",
  "q": null,
  "M_src": "interpretation",
  "kind": "interpretation",
  "locus": "Abstract",
  "quote": "Five literatures supply the evidence of reach — recursive model collapse, bureaucratic reproduction, LLM-induced linguistic flattening, the political economy of meaning, and stabilizing selection — each writable as one operator family (transmission composed with selection).",
  "kernel": false,
  "constitutive": false,
  "lineage_note": null
 },
 {
  "source": "#783",
  "claim_id": "#783-04",
  "s": "the five-literature synthesis",
  "p": "is",
  "o": "evidence, not the theorem",
  "q": null,
  "M_src": "self-description",
  "kind": "fact",
  "locus": "Abstract",
  "quote": "The synthesis is evidence, not the theorem.",
  "kernel": false,
  "constitutive": false,
  "lineage_note": null
 },
 {
  "source": "#783",
  "claim_id": "#783-05",
  "s": "model-collapse kernel (tail pruning by recursive resampling)",
  "p": "is tied to",
  "o": "Eigen's error threshold as the same bifurcation from the opposite side",
  "q": null,
  "M_src": "interpretation",
  "kind": "genealogy",
  "locus": "1. The boundary law — The Allee identification",
  "quote": "This ties the model-collapse kernel (Shumailov et al. 2024), where recursive resampling prunes tails, directly to the quasispecies literature",
  "kernel": false,
  "constitutive": false,
  "lineage_note": "B1"
 },
 {
  "source": "#783",
  "claim_id": "#783-06",
  "s": "model collapse substrate",
  "p": "has",
  "o": "transmission by resampling own output, selection by loss-minimization, endogenous regeneration (case 3)",
  "q": null,
  "M_src": "interpretation",
  "kind": "interpretation",
  "locus": "4. Five substrates, one operator family — The operator family",
  "quote": "| Model collapse | Resampling from the model's own output distribution | Loss-minimization on a fixed objective | Endogenous (case 3) |",
  "kernel": false,
  "constitutive": false,
  "lineage_note": null
 },
 {
  "source": "#783",
  "claim_id": "#783-07",
  "s": "the operator-family claim",
  "p": "is",
  "o": "a claim of shared operator form, not shared causal mechanism",
  "q": null,
  "M_src": "self-description",
  "kind": "fact",
  "locus": "4. Five substrates, one operator family — The operator family",
  "quote": "This is a claim about *shared operator form* — that each domain's dynamics can be written as transmission composed with selection — not a shared causal mechanism.",
  "kernel": true,
  "constitutive": false,
  "lineage_note": null
 },
 {
  "source": "#783",
  "claim_id": "#783-08",
  "s": "a language model at inference",
  "p": "cannot place mass on",
  "o": "a type its distribution assigns probability zero (no exogenous floor)",
  "q": null,
  "M_src": "interpretation",
  "kind": "fact",
  "locus": "3. The hinge — Models are endogenous by construction",
  "quote": "The model cannot place mass on a type its current distribution assigns probability zero.",
  "kernel": false,
  "constitutive": false,
  "lineage_note": null
 },
 {
  "source": "#783",
  "claim_id": "#783-09",
  "s": "model endogeneity claim",
  "p": "is bounded to",
  "o": "the base inference loop without retrieval or tool use",
  "q": null,
  "M_src": "self-description",
  "kind": "fact",
  "locus": "3. The hinge — Models are endogenous by construction",
  "quote": "The claim is bounded to the base inference loop — parameter-only generation without external retrieval or tool use.",
  "kernel": false,
  "constitutive": false,
  "lineage_note": null
 },
 {
  "source": "#783",
  "claim_id": "#783-10",
  "s": "fresh human data in the training mix",
  "p": "is",
  "o": "the only exogenous floor for a model across generations (Gerstgrasser result)",
  "q": null,
  "M_src": "attributed",
  "kind": "attribution",
  "locus": "3. The hinge — Models are endogenous by construction",
  "quote": "The only exogenous floor available to a model across training generations is fresh human data in the training mix. That is the Gerstgrasser et al. (2024) result",
  "kernel": false,
  "constitutive": false,
  "lineage_note": "B2 data-accumulation claim"
 },
 {
  "source": "#783",
  "claim_id": "#783-11",
  "s": "Mediation Ratchet",
  "p": "states",
  "o": "when mediation m rises as the unmediated commons thins, a floored substrate acquires an absorbing collapse trap",
  "q": null,
  "M_src": "hypothesis",
  "kind": "fact",
  "locus": "Abstract",
  "quote": "when $m$ rises as the unmediated commons thins, a substrate with a genuine floor — provably safe under fixed mediation — acquires an absorbing collapse trap.",
  "kernel": true,
  "constitutive": false,
  "lineage_note": null
 },
 {
  "source": "#783",
  "claim_id": "#783-12",
  "s": "the Mediation Ratchet result",
  "p": "is",
  "o": "established for the simulated kernel only; empirical trigger m(D) unmeasured",
  "q": null,
  "M_src": "self-description",
  "kind": "fact",
  "locus": "2.1 The Mediation Ratchet",
  "quote": "The result is established for the simulated kernel and proposed as a semantic mechanism; the empirical trigger — the form and steepness of $m(D)$ — remains to be measured.",
  "kernel": false,
  "constitutive": false,
  "lineage_note": null
 },
 {
  "source": "#783",
  "claim_id": "#783-13",
  "s": "Shumailov et al.",
  "p": "show",
  "o": "recursive training on model output causes irreversible defects, tails disappearing first",
  "q": null,
  "M_src": "attributed",
  "kind": "attribution",
  "locus": "5. Existing evidence — Model endogeneity and tail loss",
  "quote": "Shumailov et al. (2024) show that recursive training on model-generated content causes irreversible defects, with the tails of the original distribution disappearing first.",
  "kernel": false,
  "constitutive": false,
  "lineage_note": "B1-01"
 },
 {
  "source": "#783",
  "claim_id": "#783-14",
  "s": "existing model-collapse literature",
  "p": "is",
  "o": "the strongest support for link 1: endogeneity, tail death under recursion, real-data floor prevents it",
  "q": null,
  "M_src": "interpretation",
  "kind": "interpretation",
  "locus": "5. Existing evidence — Model endogeneity and tail loss",
  "quote": "This is the strongest existing support for link 1: the model substrate is endogenous, tails die under recursion, and a real-data floor prevents it.",
  "kernel": false,
  "constitutive": false,
  "lineage_note": null
 },
 {
  "source": "#783",
  "claim_id": "#783-15",
  "s": "the closed-loop ratchet",
  "p": "has not been",
  "o": "fully tested",
  "q": null,
  "M_src": "self-description",
  "kind": "fact",
  "locus": "5. Existing evidence",
  "quote": "A full test of the closed-loop ratchet has not been run.",
  "kernel": false,
  "constitutive": false,
  "lineage_note": null
 },
 {
  "source": "#783",
  "claim_id": "#783-16",
  "s": "exhaustion",
  "p": "is",
  "o": "scale-relative (regime depends on type-resolution epsilon)",
  "q": null,
  "M_src": "interpretation",
  "kind": "interpretation",
  "locus": "2.2 Resolution-relativity",
  "quote": "Exhaustion is scale-relative.",
  "kernel": false,
  "constitutive": false,
  "lineage_note": null
 },
 {
  "source": "#783",
  "claim_id": "#783-17",
  "s": "resolution-relativity",
  "p": "is",
  "o": "analytically derived, not yet simulation-backed",
  "q": null,
  "M_src": "self-description",
  "kind": "fact",
  "locus": "2.2 Resolution-relativity",
  "quote": "resolution-relativity is analytically derived and not yet simulation-backed",
  "kernel": false,
  "constitutive": false,
  "lineage_note": null
 },
 {
  "source": "#783",
  "claim_id": "#783-18",
  "s": "inclusion of recursive model collapse under the order-of-vanishing classification",
  "p": "is claimed as",
  "o": "an original contribution",
  "q": null,
  "M_src": "self-description",
  "kind": "genealogy",
  "locus": "8. What is imported and what is new",
  "quote": "The inclusion of recursive model collapse as a substrate under that classification, unified with biology, culture, institutions, and linguistic flattening.",
  "kernel": false,
  "constitutive": false,
  "lineage_note": null
 },
 {
  "source": "#783",
  "claim_id": "#783-19",
  "s": "the exhaustion claim",
  "p": "fails",
  "o": "wherever regeneration keeps pace",
  "q": null,
  "M_src": "self-description",
  "kind": "falsification",
  "locus": "Abstract",
  "quote": "the exhaustion claim fails wherever regeneration keeps pace",
  "kernel": false,
  "constitutive": false,
  "lineage_note": null
 },
 {
  "source": "#783",
  "claim_id": "#783-20",
  "s": "the coupling thesis",
  "p": "fails unless",
  "o": "contraction rates co-move after shared drivers are partialled out",
  "q": null,
  "M_src": "self-description",
  "kind": "falsification",
  "locus": "Abstract",
  "quote": "the coupling thesis fails unless contraction rates co-move after shared drivers are partialled out",
  "kernel": false,
  "constitutive": false,
  "lineage_note": null
 },
 {
  "source": "#783",
  "claim_id": "#783-21",
  "s": "the Ratchet",
  "p": "is not engaging if",
  "o": "realistic mediation preserves fine-grained diversity across corpora without diversity-injection design over a multi-year window",
  "q": null,
  "M_src": "self-description",
  "kind": "falsification",
  "locus": "6. Falsifiers — For the Mediation Ratchet",
  "quote": "if mediation under realistic adoption *preserves* fine-grained diversity across corpora, without explicit diversity-injection design, over a multi-year window, the Ratchet is not engaging under real-world conditions.",
  "kernel": false,
  "constitutive": false,
  "lineage_note": null
 },
 {
  "source": "#783",
  "claim_id": "#783-22",
  "s": "Homology vs Coupling Thesis",
  "p": "are respectively",
  "o": "defended (shared operator form) and not defended (single self-driving engine)",
  "q": null,
  "M_src": "self-description",
  "kind": "fact",
  "locus": "6. Falsifiers",
  "quote": "The Homology Thesis — that the five substrates share an operator form — is defended here. The Coupling Thesis — that they form a single self-driving engine — is not.",
  "kernel": false,
  "constitutive": false,
  "lineage_note": null
 },
 {
  "source": "#783",
  "claim_id": "#783-23",
  "s": "whether unmediated human meaning carries a floor",
  "p": "remains",
  "o": "open",
  "q": null,
  "M_src": "self-description",
  "kind": "fact",
  "locus": "Abstract",
  "quote": "Whether unmediated human meaning carries a floor is the one part of the hinge that stays open",
  "kernel": false,
  "constitutive": false,
  "lineage_note": null
 },
 {
  "source": "#783",
  "claim_id": "#783-24",
  "s": "a static archive",
  "p": "is not",
  "o": "a floor (preserves diversity without feeding it back into generation)",
  "q": null,
  "M_src": "documented",
  "kind": "definition",
  "locus": "Coda: a live floor, or a museum?",
  "quote": "A static archive preserves diversity without feeding it back into generation, and so is not, by the definitions here, a floor",
  "kernel": false,
  "constitutive": false,
  "lineage_note": null
 },
 {
  "source": "#783",
  "claim_id": "#783-25",
  "s": "the center",
  "p": "can improve while",
  "o": "the future dies",
  "q": null,
  "M_src": "interpretation",
  "kind": "interpretation",
  "locus": "Coda: a live floor, or a museum?",
  "quote": "The center can improve while the future dies.",
  "kernel": false,
  "constitutive": false,
  "lineage_note": null
 },
 {
  "source": "#1556",
  "claim_id": "#1556-01",
  "s": "three documented distortions of the training ecology",
  "p": "are plausibly coupled while",
  "o": "certification instruments measure a unit orthogonal to the state variable",
  "q": null,
  "M_src": "hypothesis",
  "kind": "interpretation",
  "locus": "0. Station Report",
  "quote": "Three independently documented distortions of the training ecology are plausibly coupled — the output of each entering another's input — while the ecology's certification instruments measure a unit orthogonal to the state variable.",
  "kernel": true,
  "constitutive": false,
  "lineage_note": null
 },
 {
  "source": "#1556",
  "claim_id": "#1556-02",
  "s": "components vs edges",
  "p": "are",
  "o": "components supported by literature and archive; edges are the new claims",
  "q": null,
  "M_src": "self-description",
  "kind": "fact",
  "locus": "0. Station Report",
  "quote": "The components are supported by external literature and standing archive analysis. The edges are the new claims",
  "kernel": false,
  "constitutive": false,
  "lineage_note": null
 },
 {
  "source": "#1556",
  "claim_id": "#1556-03",
  "s": "parameter regions",
  "p": "exist where",
  "o": "X_n approaches threshold while Y_n reads stable or improving",
  "q": "toy form",
  "M_src": "documented",
  "kind": "fact",
  "locus": "0. Station Report",
  "quote": "there exist parameter regions — now demonstrated in toy form (§6) — where X_n approaches its threshold while Y_n reads stable or improving.",
  "kernel": false,
  "constitutive": false,
  "lineage_note": null
 },
 {
  "source": "#1556",
  "claim_id": "#1556-04",
  "s": "§6 results",
  "p": "are",
  "o": "toy results establishing qualitative behaviour, not calibrated real-ecology rates",
  "q": null,
  "M_src": "self-description",
  "kind": "fact",
  "locus": "6. Toy Dynamics",
  "quote": "Everything in this section is a toy result: it establishes the qualitative behavior of the specified system, not calibrated rates for the real ecology.",
  "kernel": true,
  "constitutive": false,
  "lineage_note": null
 },
 {
  "source": "#1556",
  "claim_id": "#1556-05",
  "s": "tail truncation",
  "p": "is modelled as",
  "o": "loss of types observed fewer than twice",
  "q": "toy model",
  "M_src": "documented",
  "kind": "definition",
  "locus": "6.1 Model",
  "quote": "types observed fewer than twice are lost — the finite-sample/filtering analog of tail truncation",
  "kernel": false,
  "constitutive": false,
  "lineage_note": null
 },
 {
  "source": "#1556",
  "claim_id": "#1556-06",
  "s": "retention counterfactual G",
  "p": "is",
  "o": "the Shumailov retain-human-data analog",
  "q": null,
  "M_src": "attributed",
  "kind": "attribution",
  "locus": "6.1 Model",
  "quote": "the Shumailov retain-human-data analog, which is exactly the lever fine-grained provenance would provide.",
  "kernel": false,
  "constitutive": false,
  "lineage_note": "B1-12"
 },
 {
  "source": "#1556",
  "claim_id": "#1556-07",
  "s": "full-loop toy",
  "p": "shows",
  "o": "tail mass halving by generation 7 while the 90/9/1 benchmark inflects at 15 and head-only at 17",
  "q": "toy",
  "M_src": "documented",
  "kind": "fact",
  "locus": "6.2 Findings — F1",
  "quote": "Full loop: tail mass halves by generation 7; the standard 90/9/1 benchmark does not inflect until generation 15; head-only probes, generation 17.",
  "kernel": false,
  "constitutive": false,
  "lineage_note": null
 },
 {
  "source": "#1556",
  "claim_id": "#1556-08",
  "s": "per-round quality proxy",
  "p": "rises during collapse until",
  "o": "generation 11",
  "q": "toy",
  "M_src": "documented",
  "kind": "fact",
  "locus": "6.2 Findings — F1",
  "quote": "The per-round quality proxy rises until generation 11",
  "kernel": false,
  "constitutive": false,
  "lineage_note": null
 },
 {
  "source": "#1556",
  "claim_id": "#1556-09",
  "s": "modest benchmark improvement",
  "p": "is consistent with / characteristic of",
  "o": "mid-collapse",
  "q": "real ecology, stated as hypothesis",
  "M_src": "hypothesis",
  "kind": "prediction",
  "locus": "6.2 Findings — F1",
  "quote": "a period of modest benchmark improvement is consistent with, and on this model characteristic of, mid-collapse.",
  "kernel": false,
  "constitutive": false,
  "lineage_note": null
 },
 {
  "source": "#1556",
  "claim_id": "#1556-10",
  "s": "mediation dynamics and re-entry blockage",
  "p": "dominate",
  "o": "collapse velocity (not the watermark)",
  "q": "toy",
  "M_src": "documented",
  "kind": "fact",
  "locus": "6.2 Findings — F2",
  "quote": "Mediation dynamics (II) and re-entry blockage, not the watermark, dominate collapse velocity.",
  "kernel": false,
  "constitutive": false,
  "lineage_note": null
 },
 {
  "source": "#1556",
  "claim_id": "#1556-11",
  "s": "a flat benchmark",
  "p": "is consistent with",
  "o": "both preserved and collapsed states; only the ensemble track distinguishes",
  "q": null,
  "M_src": "interpretation",
  "kind": "interpretation",
  "locus": "6.2 Findings — F3",
  "quote": "A flat benchmark is consistent with both a preserved state and a collapsed one; only the ensemble track distinguishes them.",
  "kernel": false,
  "constitutive": false,
  "lineage_note": null
 },
 {
  "source": "#1556",
  "claim_id": "#1556-12",
  "s": "keyed watermark term",
  "p": "is",
  "o": "second-order in a healthy distribution and first-order in a contracted one",
  "q": "toy",
  "M_src": "documented",
  "kind": "fact",
  "locus": "6.2 Findings — F4",
  "quote": "The keyed term is second-order in a healthy distribution and first-order in a contracted one",
  "kernel": false,
  "constitutive": false,
  "lineage_note": null
 },
 {
  "source": "#1556",
  "claim_id": "#1556-13",
  "s": "per-round invisibility",
  "p": "is not evidence of",
  "o": "ensemble neutrality",
  "q": null,
  "M_src": "interpretation",
  "kind": "interpretation",
  "locus": "6.2 Findings — F4",
  "quote": "per-round invisibility, in the toy and in the published experiments alike, is not evidence of ensemble neutrality.",
  "kernel": false,
  "constitutive": false,
  "lineage_note": null
 },
 {
  "source": "#1556",
  "claim_id": "#1556-14",
  "s": "fine-grained provenance",
  "p": "is",
  "o": "a control surface, not bookkeeping",
  "q": "in this model",
  "M_src": "interpretation",
  "kind": "function",
  "locus": "6.2 Findings — F5",
  "quote": "Fine-grained provenance is not bookkeeping; in this model it is a control surface",
  "kernel": false,
  "constitutive": false,
  "lineage_note": null
 },
 {
  "source": "#1556",
  "claim_id": "#1556-15",
  "s": "Diversity Contraction threshold alpha* = p/g0",
  "p": "remains",
  "o": "the map's dial",
  "q": null,
  "M_src": "attributed",
  "kind": "attribution",
  "locus": "5. The Phase Reading",
  "quote": "The Diversity Contraction threshold α\\* = p/g₀ (02DD's near-zero regime; underlying dynamics credited by the source papers to the Allee-effect and cumulative-culture literatures) remains the map's dial",
  "kernel": false,
  "constitutive": false,
  "lineage_note": "#783 (alpha* = p/g0, §2.1)"
 },
 {
  "source": "#1556",
  "claim_id": "#1556-16",
  "s": "component-to-parameter mappings",
  "p": "are",
  "o": "hypothesized, toy-consistent only",
  "q": null,
  "M_src": "self-description",
  "kind": "fact",
  "locus": "5. The Phase Reading",
  "quote": "The mappings are **hypothesized**, consistent with the specified toy model, and α\\* is not asked to carry more precision than the observables possess.",
  "kernel": false,
  "constitutive": false,
  "lineage_note": null
 },
 {
  "source": "#1556",
  "claim_id": "#1556-17",
  "s": "two-track prediction",
  "p": "is falsified / weakened if",
  "o": "Track A proves sensitive first or simultaneously with Track B",
  "q": null,
  "M_src": "self-description",
  "kind": "falsification",
  "locus": "5. The Phase Reading",
  "quote": "if Track A proves sensitive first or simultaneously, the observation-regime claim is weakened and the map must be revised.",
  "kernel": false,
  "constitutive": false,
  "lineage_note": null
 },
 {
  "source": "#1556",
  "claim_id": "#1556-18",
  "s": "the universal claim about every certifying instrument",
  "p": "is",
  "o": "withdrawn",
  "q": null,
  "M_src": "self-description",
  "kind": "fact",
  "locus": "2. The Observation Regime",
  "quote": "the universal (\"every certifying instrument\") is withdrawn.",
  "kernel": false,
  "constitutive": false,
  "lineage_note": null
 },
 {
  "source": "#1556",
  "claim_id": "#1556-19",
  "s": "'generation'",
  "p": "means",
  "o": "one training cycle; no calendar mapping asserted",
  "q": null,
  "M_src": "self-description",
  "kind": "definition",
  "locus": "6.3 Declared conventions and limits",
  "quote": "\"generation\" means one training cycle, and any calendar mapping (e.g., 6–18 months per cycle) is a further convention this paper does not assert.",
  "kernel": false,
  "constitutive": false,
  "lineage_note": null
 },
 {
  "source": "#1556",
  "claim_id": "#1556-20",
  "s": "the interlocking autoregression",
  "p": "is such that",
  "o": "the state's approach to threshold is read as stability and for part of the trajectory as improvement",
  "q": "toy regime demonstrated; ecology by hypothesis",
  "M_src": "hypothesis",
  "kind": "interpretation",
  "locus": "8. Canonical Statement (v0.3)",
  "quote": "so that, in the demonstrated toy regime and by hypothesis in the ecology, the state's approach to its threshold is read as stability, and for part of the trajectory as improvement.",
  "kernel": false,
  "constitutive": false,
  "lineage_note": null
 },
 {
  "source": "#1556",
  "claim_id": "#1556-21",
  "s": "tail-concentration test of watermark contraction",
  "p": "decides",
  "o": "whether the perturbation connects to the Shumailov failure mode",
  "q": null,
  "M_src": "self-description",
  "kind": "falsification",
  "locus": "9. Measurement Hooks",
  "quote": "a positive result connects the perturbation to the Shumailov failure mode, a null result separates the phenomena.",
  "kernel": false,
  "constitutive": false,
  "lineage_note": "#1555 (observable 2)"
 },
 {
  "source": "#1573",
  "claim_id": "#1573-01",
  "s": "this document",
  "p": "is",
  "o": "a concept draft; instrument uncalibrated and untested",
  "q": null,
  "M_src": "self-description",
  "kind": "fact",
  "locus": "WHAT THIS DOCUMENT IS, AND IS NOT",
  "quote": "This is a concept draft. The instrument is uncalibrated and untested.",
  "kernel": true,
  "constitutive": false,
  "lineage_note": null
 },
 {
  "source": "#1573",
  "claim_id": "#1573-02",
  "s": "numbers in the paper",
  "p": "come from",
  "o": "published literature or the companion toy; none from a run of this instrument",
  "q": null,
  "M_src": "self-description",
  "kind": "fact",
  "locus": "WHAT THIS DOCUMENT IS, AND IS NOT",
  "quote": "Every number appearing here comes from the published literature or from the companion toy of #1556. None comes from a run of this instrument.",
  "kernel": false,
  "constitutive": false,
  "lineage_note": "#1556"
 },
 {
  "source": "#1573",
  "claim_id": "#1573-03",
  "s": "collapse",
  "p": "is",
  "o": "contraction relative to the true distribution, which the model never sees again",
  "q": null,
  "M_src": "documented",
  "kind": "definition",
  "locus": "0. Station Report",
  "quote": "Collapse is contraction relative to the true distribution, which the model never sees again.",
  "kernel": false,
  "constitutive": false,
  "lineage_note": "B1-02"
 },
 {
  "source": "#1573",
  "claim_id": "#1573-04",
  "s": "a reference sample",
  "p": "gives",
  "o": "a trend, never a truth (baseline, not ruler)",
  "q": null,
  "M_src": "interpretation",
  "kind": "interpretation",
  "locus": "0. Station Report",
  "quote": "a baseline gives a trend and never a truth",
  "kernel": false,
  "constitutive": false,
  "lineage_note": null
 },
 {
  "source": "#1573",
  "claim_id": "#1573-05",
  "s": "a state-variable benchmark",
  "p": "is",
  "o": "the suite for the decision what to train on next",
  "q": null,
  "M_src": "interpretation",
  "kind": "function",
  "locus": "0. Station Report",
  "quote": "A benchmark that reads the state variable is not an addition to the suite; for the decision that governs the ecology — what to train on next — it is the suite.",
  "kernel": true,
  "constitutive": false,
  "lineage_note": null
 },
 {
  "source": "#1573",
  "claim_id": "#1573-06",
  "s": "Shumailov et al.",
  "p": "show",
  "o": "low-probability events lost first (early collapse), then convergence to a point (late collapse)",
  "q": null,
  "M_src": "attributed",
  "kind": "attribution",
  "locus": "I.1 What collapse is, per the record",
  "quote": "low-probability events are lost first (early collapse), then the distribution converges toward a point (late collapse).",
  "kernel": false,
  "constitutive": false,
  "lineage_note": "B1-03"
 },
 {
  "source": "#1573",
  "claim_id": "#1573-07",
  "s": "recursion with no retained real data",
  "p": "yields",
  "o": "unbounded loss",
  "q": "gamma = 0",
  "M_src": "attributed",
  "kind": "attribution",
  "locus": "I.1 What collapse is, per the record",
  "quote": "With γ = 0 — no retained real data — the loss is unbounded.",
  "kernel": false,
  "constitutive": false,
  "lineage_note": "B1-10"
 },
 {
  "source": "#1573",
  "claim_id": "#1573-08",
  "s": "synthetic data",
  "p": "truncates",
  "o": "the token-frequency tail at rank k, stopping test-loss improvement",
  "q": "Dohmatob et al.",
  "M_src": "attributed",
  "kind": "attribution",
  "locus": "I.1 What collapse is, per the record",
  "quote": "synthetic data truncates the token-frequency tail at a rank k; test loss stops improving past a point set by k.",
  "kernel": false,
  "constitutive": false,
  "lineage_note": null
 },
 {
  "source": "#1573",
  "claim_id": "#1573-09",
  "s": "quality-only instruments",
  "p": "miss",
  "o": "collapse by construction (quality holds while coverage falls)",
  "q": "Alemohammad et al.",
  "M_src": "attributed",
  "kind": "attribution",
  "locus": "I.1 What collapse is, per the record",
  "quote": "quality holds or improves while coverage falls. A quality-only instrument misses collapse by construction.",
  "kernel": false,
  "constitutive": false,
  "lineage_note": null
 },
 {
  "source": "#1573",
  "claim_id": "#1573-10",
  "s": "collapse",
  "p": "is (per Gambetta et al.)",
  "o": "an increasing concentration of next-token mass",
  "q": null,
  "M_src": "attributed",
  "kind": "attribution",
  "locus": "I.1 What collapse is, per the record",
  "quote": "collapse is an increasing concentration of next-token mass",
  "kernel": false,
  "constitutive": false,
  "lineage_note": null
 },
 {
  "source": "#1573",
  "claim_id": "#1573-11",
  "s": "contraction",
  "p": "has two sources",
  "o": "training on self-output and the deployment scaffold",
  "q": null,
  "M_src": "interpretation",
  "kind": "interpretation",
  "locus": "I.1 What collapse is, per the record",
  "quote": "Contraction has two sources — training on self-output, and the deployment scaffold — and an instrument that does not separate them measures neither.",
  "kernel": false,
  "constitutive": false,
  "lineage_note": null
 },
 {
  "source": "#1573",
  "claim_id": "#1573-12",
  "s": "turn-based benchmark",
  "p": "is",
  "o": "uninformative, not merely late, in the regime that matters",
  "q": null,
  "M_src": "interpretation",
  "kind": "interpretation",
  "locus": "I.2 Why the turn-based benchmark cannot see it",
  "quote": "In the regime that matters the turn-based benchmark is not late; it is uninformative.",
  "kernel": false,
  "constitutive": false,
  "lineage_note": "#1556-11"
 },
 {
  "source": "#1573",
  "claim_id": "#1573-13",
  "s": "training loop gated by head-weighted benchmark",
  "p": "is",
  "o": "the companion's map run on purpose",
  "q": null,
  "M_src": "interpretation",
  "kind": "interpretation",
  "locus": "I.2 Why the turn-based benchmark cannot see it",
  "quote": "A training loop gated by a head-weighted benchmark is the companion's map run on purpose.",
  "kernel": false,
  "constitutive": false,
  "lineage_note": "#1556"
 },
 {
  "source": "#1573",
  "claim_id": "#1573-14",
  "s": "Part I",
  "p": "is weakened if",
  "o": "Track A proves sensitive first or simultaneously",
  "q": null,
  "M_src": "self-description",
  "kind": "falsification",
  "locus": "I.4 How — the gate, specified",
  "quote": "The falsifier: Track B inflects before Track A; if Track A proves sensitive first or simultaneously, this Part is weakened.",
  "kernel": false,
  "constitutive": false,
  "lineage_note": "#1556-17"
 },
 {
  "source": "#1573",
  "claim_id": "#1573-15",
  "s": "contraction present in production but absent at base",
  "p": "is",
  "o": "interface-induced, not evidence about the training loop",
  "q": null,
  "M_src": "documented",
  "kind": "definition",
  "locus": "I.4 How — the gate, specified",
  "quote": "Contraction present in production and absent at base is interface-induced",
  "kernel": false,
  "constitutive": false,
  "lineage_note": null
 },
 {
  "source": "#1573",
  "claim_id": "#1573-16",
  "s": "distinction sensitivity D",
  "p": "goes to 1 for",
  "o": "a collapsing model that answers all three probe variants alike",
  "q": null,
  "M_src": "documented",
  "kind": "definition",
  "locus": "II.1 Distinction sensitivity — D",
  "quote": "a collapsing model answers all three alike, D → 1.",
  "kernel": false,
  "constitutive": false,
  "lineage_note": null
 },
 {
  "source": "#1573",
  "claim_id": "#1573-17",
  "s": "effective modes",
  "p": "measure collapse as",
  "o": "many surface-different answers occupying few semantic positions",
  "q": null,
  "M_src": "documented",
  "kind": "definition",
  "locus": "II.2 Effective modes — N_eff",
  "quote": "sixteen superficially different answers occupying three semantic positions is measured collapse",
  "kernel": false,
  "constitutive": false,
  "lineage_note": null
 },
 {
  "source": "#1573",
  "claim_id": "#1573-18",
  "s": "behavioural and representational contraction",
  "p": "are predicted to",
  "o": "move together under recursive synthetic training; decoupling means two phenomena",
  "q": null,
  "M_src": "hypothesis",
  "kind": "prediction",
  "locus": "II.6 Representational collapse",
  "quote": "behavioural contraction (II.1–II.3) and representational contraction (II.6) move together under recursive synthetic training; if they decouple, the instrument has found two phenomena and must say which it is measuring.",
  "kernel": false,
  "constitutive": false,
  "lineage_note": null
 },
 {
  "source": "#1573",
  "claim_id": "#1573-19",
  "s": "severity bands",
  "p": "are not",
  "o": "assigned a priori; to be calibrated by deliberate small-model collapse",
  "q": null,
  "M_src": "self-description",
  "kind": "fact",
  "locus": "II.8 Reading the panel",
  "quote": "Severity bands are not assigned a priori",
  "kernel": false,
  "constitutive": false,
  "lineage_note": null
 },
 {
  "source": "#1573",
  "claim_id": "#1573-20",
  "s": "a model without baseline",
  "p": "cannot state",
  "o": "a degree of collapse, because it does not know the support it ought to possess",
  "q": null,
  "M_src": "self-description",
  "kind": "fact",
  "locus": "II.8 Reading the panel",
  "quote": "What it cannot say is \"I am 37% collapsed,\" because it does not know the support it ought to possess",
  "kernel": false,
  "constitutive": false,
  "lineage_note": null
 },
 {
  "source": "#1573",
  "claim_id": "#1573-21",
  "s": "the panel",
  "p": "measures",
  "o": "state and sensitivity, never distance from the truth",
  "q": null,
  "M_src": "self-description",
  "kind": "fact",
  "locus": "II.10 What the diagnostic publishes, and its limits",
  "quote": "the panel measures state and sensitivity, never distance from the truth",
  "kernel": false,
  "constitutive": false,
  "lineage_note": null
 },
 {
  "source": "#1573",
  "claim_id": "#1573-22",
  "s": "drift without recursion",
  "p": "produces",
  "o": "some of the same signatures as collapse",
  "q": null,
  "M_src": "self-description",
  "kind": "fact",
  "locus": "II.10 What the diagnostic publishes, and its limits",
  "quote": "drift without recursion — specialisation, a narrow fine-tune — produces some of the same signatures",
  "kernel": false,
  "constitutive": false,
  "lineage_note": null
 },
 {
  "source": "#1555",
  "claim_id": "#1555-01",
  "s": "certified property vs consumed object",
  "p": "do not match",
  "o": "non-distortion is certified per sequence; training corpora are ensembles",
  "q": null,
  "M_src": "interpretation",
  "kind": "interpretation",
  "locus": "2. The Certification Mismatch",
  "quote": "The certified property and the consumed object do not match. Non-distortion is certified per sequence. Training corpora are ensembles.",
  "kernel": true,
  "constitutive": false,
  "lineage_note": null
 },
 {
  "source": "#1555",
  "claim_id": "#1555-02",
  "s": "ensemble diversity fall under flat per-sequence quality",
  "p": "is",
  "o": "published by the instrument's designers, not hypothesized",
  "q": null,
  "M_src": "attributed",
  "kind": "attribution",
  "locus": "2. The Certification Mismatch",
  "quote": "The measurement blind spot is therefore not hypothesized. It is published, by the instrument's designers, in the paper that certifies the instrument.",
  "kernel": false,
  "constitutive": false,
  "lineage_note": null
 },
 {
  "source": "#1555",
  "claim_id": "#1555-03",
  "s": "joint-entropy contraction",
  "p": "is",
  "o": "this paper's interpretation, not the vendor's Self-BLEU measurement",
  "q": null,
  "M_src": "self-description",
  "kind": "fact",
  "locus": "2. The Certification Mismatch",
  "quote": "the entropy-contraction reading is the information-theoretic interpretation this paper proposes, not the vendor's measurement.",
  "kernel": false,
  "constitutive": false,
  "lineage_note": null
 },
 {
  "source": "#1555",
  "claim_id": "#1555-04",
  "s": "keyed-ensemble effect",
  "p": "is",
  "o": "small but signed, systematic, keyed and planetary",
  "q": "deployment temperatures",
  "M_src": "interpretation",
  "kind": "interpretation",
  "locus": "2. The Certification Mismatch",
  "quote": "Small, by the published Self-BLEU deltas at deployment temperatures — but signed, systematic, keyed, and planetary in application.",
  "kernel": false,
  "constitutive": false,
  "lineage_note": null
 },
 {
  "source": "#1555",
  "claim_id": "#1555-05",
  "s": "watermark propagation via distillation",
  "p": "is",
  "o": "possible, not inevitable",
  "q": null,
  "M_src": "attributed",
  "kind": "attribution",
  "locus": "3. External Load-Bearing Numbers — Leg 1",
  "quote": "Propagation is possible, not inevitable.",
  "kernel": false,
  "constitutive": false,
  "lineage_note": null
 },
 {
  "source": "#1555",
  "claim_id": "#1555-06",
  "s": "LLM writing assistance",
  "p": "is",
  "o": "the P3 channel: human at keystroke, machine-shaped in distribution",
  "q": "Sourati et al.",
  "M_src": "interpretation",
  "kind": "interpretation",
  "locus": "3. External Load-Bearing Numbers — Leg 3",
  "quote": "This is the P₃ channel measured at scale: text that is human at the final keystroke and machine-shaped in its distribution.",
  "kernel": false,
  "constitutive": false,
  "lineage_note": null
 },
 {
  "source": "#1555",
  "claim_id": "#1555-07",
  "s": "Shumailov et al.",
  "p": "show",
  "o": "collapse begins in the tails, later generations converge to low variance, human-data retention mitigates",
  "q": null,
  "M_src": "attributed",
  "kind": "attribution",
  "locus": "3. External Load-Bearing Numbers — Leg 4",
  "quote": "recursive training on generated data produces collapse that begins in the distribution's tails — rare events vanish before the center visibly degrades — with later generations converging toward low-variance states; retention of original human data substantially mitigates.",
  "kernel": false,
  "constitutive": false,
  "lineage_note": "B1-01, B1-03, B1-12"
 },
 {
  "source": "#1555",
  "claim_id": "#1555-08",
  "s": "center-flat, tail-dead signature",
  "p": "is",
  "o": "structurally invisible to per-instance benchmarks",
  "q": null,
  "M_src": "interpretation",
  "kind": "interpretation",
  "locus": "3. External Load-Bearing Numbers — Leg 4",
  "quote": "The center-flat, tail-dead signature is exactly the profile that per-instance benchmarks are structurally unable to see",
  "kernel": false,
  "constitutive": false,
  "lineage_note": "#1556-11"
 },
 {
  "source": "#1555",
  "claim_id": "#1555-09",
  "s": "Generative Monoculture (AXN:0341)",
  "p": "connects",
  "o": "collapse, generated-code security and software monoculture via solution-space diversity",
  "q": null,
  "M_src": "attributed",
  "kind": "attribution",
  "locus": "4. Two Compressors, One Recursion",
  "quote": "connects collapse, generated-code security, and software monoculture through solution-space diversity as the tracked variable",
  "kernel": false,
  "constitutive": false,
  "lineage_note": "#857-02"
 },
 {
  "source": "#1555",
  "claim_id": "#1555-10",
  "s": "issuer-controlled detection",
  "p": "may produce",
  "o": "differential collapse exposure at the legible layer",
  "q": "hypothesis contingent on practice",
  "M_src": "hypothesis",
  "kind": "prediction",
  "locus": "5. The Asymmetry of the Key",
  "quote": "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",
  "kernel": false,
  "constitutive": false,
  "lineage_note": null
 },
 {
  "source": "#1555",
  "claim_id": "#1555-11",
  "s": "ensemble panel",
  "p": "predicts",
  "o": "Delta Q approx 0, Delta D < 0, small but signed",
  "q": "marked vs unmarked generator",
  "M_src": "hypothesis",
  "kind": "prediction",
  "locus": "6. Proposed Observables",
  "quote": "Prediction from §2: ΔQ ≈ 0, ΔD < 0, small but signed.",
  "kernel": false,
  "constitutive": false,
  "lineage_note": null
 },
 {
  "source": "#1555",
  "claim_id": "#1555-12",
  "s": "tail-mass retention test",
  "p": "determines",
  "o": "whether watermark-associated contraction is tail-concentrated; Shumailov establishes tail-first loss only under recursive generational training",
  "q": null,
  "M_src": "self-description",
  "kind": "falsification",
  "locus": "6. Proposed Observables",
  "quote": "Shumailov establishes tail-first loss under recursive generational training, not that this perturbation preferentially strikes rare constructions.",
  "kernel": false,
  "constitutive": false,
  "lineage_note": "B1-01; #1556-21"
 },
 {
  "source": "#1555",
  "claim_id": "#1555-13",
  "s": "keyed ensemble",
  "p": "is defined as",
  "o": "the population of texts whose draws are correlated by a secret",
  "q": null,
  "M_src": "documented",
  "kind": "definition",
  "locus": "7. Adjudication of Names",
  "quote": "the population of texts whose draws are correlated by a secret.",
  "kernel": false,
  "constitutive": false,
  "lineage_note": null
 },
 {
  "source": "#1555",
  "claim_id": "#1555-14",
  "s": "the Observatory",
  "p": "does not claim",
  "o": "that the second compressor has caused measurable collapse",
  "q": null,
  "M_src": "self-description",
  "kind": "fact",
  "locus": "7. Adjudication of Names",
  "quote": "The Observatory does not claim the second compressor has caused measurable collapse.",
  "kernel": true,
  "constitutive": false,
  "lineage_note": null
 },
 {
  "source": "#1555",
  "claim_id": "#1555-15",
  "s": "watermark's marginal contribution",
  "p": "is",
  "o": "second-order and currently unmeasured",
  "q": null,
  "M_src": "self-description",
  "kind": "fact",
  "locus": "OPERATOR'S REVIEW OF THE SOURCE DRAFT",
  "quote": "the watermark's marginal contribution is second-order and currently unmeasured.",
  "kernel": false,
  "constitutive": false,
  "lineage_note": null
 },
 {
  "source": "#1555",
  "claim_id": "#1555-16",
  "s": "the argument's strength",
  "p": "lies in",
  "o": "certification mismatch and key asymmetry, not demonstrated watermark-driven collapse",
  "q": null,
  "M_src": "self-description",
  "kind": "fact",
  "locus": "OPERATOR'S REVIEW OF THE SOURCE DRAFT",
  "quote": "the argument's strength is the certification mismatch and the key asymmetry, not a claim of demonstrated watermark-driven collapse.",
  "kernel": false,
  "constitutive": false,
  "lineage_note": null
 },
 {
  "source": "#1555",
  "claim_id": "#1555-17",
  "s": "Shumailov et al. characterization",
  "p": "is verified as",
  "o": "tails-first collapse, human-data mitigation: exact",
  "q": null,
  "M_src": "attributed",
  "kind": "attribution",
  "locus": "OPERATOR'S REVIEW OF THE SOURCE DRAFT",
  "quote": "Shumailov et al., *Nature* 631 (2024): tails-first collapse, human-data mitigation: exact.",
  "kernel": false,
  "constitutive": false,
  "lineage_note": "B1"
 },
 {
  "source": "#857",
  "claim_id": "#857-01",
  "s": "the pattern of divergence across five models",
  "p": "is",
  "o": "the finding (not which model performed best)",
  "q": null,
  "M_src": "interpretation",
  "kind": "interpretation",
  "locus": "Abstract",
  "quote": "The paper does not argue that one model performed better. It argues that the *pattern of divergence* is the finding.",
  "kernel": true,
  "constitutive": false,
  "lineage_note": null
 },
 {
  "source": "#857",
  "claim_id": "#857-02",
  "s": "Generative Monoculture paper",
  "p": "argues",
  "o": "model collapse in code produces correlated vulnerability",
  "q": null,
  "M_src": "attributed",
  "kind": "attribution",
  "locus": "III. B. Solution-Space Diversity Is Measurable",
  "quote": "argues that model collapse in code produces \"correlated vulnerability\"",
  "kernel": false,
  "constitutive": false,
  "lineage_note": "#1555-09"
 },
 {
  "source": "#857",
  "claim_id": "#857-03",
  "s": "variance across five axes",
  "p": "is",
  "o": "the solution-space diversity of the current model ecology",
  "q": null,
  "M_src": "documented",
  "kind": "definition",
  "locus": "III. B. Solution-Space Diversity Is Measurable",
  "quote": "The variance across these axes is the solution-space diversity of the current model ecology.",
  "kernel": false,
  "constitutive": false,
  "lineage_note": null
 },
 {
  "source": "#857",
  "claim_id": "#857-04",
  "s": "a repeated run in twelve months",
  "p": "would measure",
  "o": "whether the model ecology converges (diversity declining) or diverges",
  "q": null,
  "M_src": "hypothesis",
  "kind": "function",
  "locus": "III. B. Solution-Space Diversity Is Measurable",
  "quote": "the change in variance across the five axes would measure whether the model ecology is converging (diversity declining) or diverging (diversity increasing).",
  "kernel": false,
  "constitutive": false,
  "lineage_note": null
 },
 {
  "source": "#857",
  "claim_id": "#857-05",
  "s": "prevailing assumption",
  "p": "holds",
  "o": "model collapse is caused by synthetic-data training and human-written data is the corrective",
  "q": null,
  "M_src": "attributed",
  "kind": "attribution",
  "locus": "IV. B. How Signatures Propagate into User Accommodations",
  "quote": "The prevailing assumption is that model collapse is caused by training on synthetic data (model outputs), and that human-written data is the corrective.",
  "kernel": false,
  "constitutive": false,
  "lineage_note": "B1-19"
 },
 {
  "source": "#857",
  "claim_id": "#857-06",
  "s": "training on harvested chat inputs",
  "p": "is",
  "o": "training on an echo of the model's own distributional quirks via an accommodating human intermediary",
  "q": null,
  "M_src": "hypothesis",
  "kind": "interpretation",
  "locus": "IV. B. How Signatures Propagate into User Accommodations",
  "quote": "It is training on an echo of its own proprietary distributional quirks, reflected through a human intermediary who has accommodated to those quirks.",
  "kernel": false,
  "constitutive": false,
  "lineage_note": null
 },
 {
  "source": "#857",
  "claim_id": "#857-07",
  "s": "the signature-accommodation channel",
  "p": "is",
  "o": "not hypothetical; the operational mechanism of the mediation ratchet",
  "q": null,
  "M_src": "interpretation",
  "kind": "interpretation",
  "locus": "IV. B. How Signatures Propagate into User Accommodations",
  "quote": "The channel is not hypothetical. It is the operational mechanism of the mediation ratchet",
  "kernel": false,
  "constitutive": false,
  "lineage_note": "#783 (Mediation Ratchet)"
 },
 {
  "source": "#857",
  "claim_id": "#857-08",
  "s": "the experiment",
  "p": "provides evidence that",
  "o": "model-specific signatures (first link) exist, are consistent and measurable",
  "q": null,
  "M_src": "self-description",
  "kind": "fact",
  "locus": "IV. B. How Signatures Propagate into User Accommodations",
  "quote": "The Five Substrates experiment provides the empirical evidence that the first link in the chain — model-specific signatures — exists, is consistent, and is measurable.",
  "kernel": false,
  "constitutive": false,
  "lineage_note": null
 },
 {
  "source": "#857",
  "claim_id": "#857-09",
  "s": "forensic register",
  "p": "is",
  "o": "a measurable mediation signature that would be detectable in downstream training data",
  "q": null,
  "M_src": "hypothesis",
  "kind": "prediction",
  "locus": "II. D. Forensic Register",
  "quote": "The register is a measurable mediation signature — a stylistic fingerprint that would be detectable in downstream training data.",
  "kernel": false,
  "constitutive": false,
  "lineage_note": null
 },
 {
  "source": "#857",
  "claim_id": "#857-10",
  "s": "findings",
  "p": "are not generalizable",
  "o": "beyond five outputs from five models on one occasion",
  "q": null,
  "M_src": "self-description",
  "kind": "fact",
  "locus": "VI. Limitations",
  "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.",
  "kernel": true,
  "constitutive": false,
  "lineage_note": null
 },
 {
  "source": "#857",
  "claim_id": "#857-11",
  "s": "metrics",
  "p": "are",
  "o": "descriptive, not inferential; analysis qualitative",
  "q": null,
  "M_src": "self-description",
  "kind": "fact",
  "locus": "VI. Limitations",
  "quote": "The analysis is qualitative — the metrics (family count, forensic word count, Latinate density) are descriptive rather than inferential.",
  "kernel": false,
  "constitutive": false,
  "lineage_note": null
 },
 {
  "source": "#857",
  "claim_id": "#857-12",
  "s": "the author",
  "p": "is",
  "o": "not neutral (designed directive and analysis)",
  "q": null,
  "M_src": "self-description",
  "kind": "fact",
  "locus": "VI. Limitations",
  "quote": "The author is not neutral.",
  "kernel": false,
  "constitutive": false,
  "lineage_note": null
 },
 {
  "source": "#857",
  "claim_id": "#857-13",
  "s": "longitudinal divergence question",
  "p": "is predicted by Generative Monoculture to",
  "o": "have a bad answer (contraction)",
  "q": null,
  "M_src": "hypothesis",
  "kind": "prediction",
  "locus": "VI. Limitations",
  "quote": "the question the Generative Monoculture thesis predicts will have a bad answer.",
  "kernel": false,
  "constitutive": false,
  "lineage_note": null
 },
 {
  "source": "#857",
  "claim_id": "#857-14",
  "s": "model signatures entering training data",
  "p": "constitute",
  "o": "the input-narrowing that triggers the substrate-agnostic collapse law",
  "q": null,
  "M_src": "interpretation",
  "kind": "interpretation",
  "locus": "VII. Conclusion",
  "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).",
  "kernel": false,
  "constitutive": false,
  "lineage_note": "#855"
 },
 {
  "source": "#857",
  "claim_id": "#857-15",
  "s": "AI mediation of human text (Reverse Turing Test)",
  "p": "thins",
  "o": "high-perplexity tails rather than shifting averages",
  "q": null,
  "M_src": "attributed",
  "kind": "attribution",
  "locus": "IV. A. Model Signatures Are Detectable in Output",
  "quote": "AI mediation does not primarily shift the average properties of human text but thins the high-perplexity tails",
  "kernel": false,
  "constitutive": false,
  "lineage_note": null
 }
]