Model monoculture: SSDI in code, entity substitution and reflexive foreclosure

These claims predict that recursive synthetic saturation narrows model outputs: Generative Monoculture (#199) forecasts a falling SSDI and rising cross-codebase vulnerability correlation while pass rates hold, and the Wolf Boy paper states a substrate-agnostic law of capacity loss under contracting input diversity. Reflexive Foreclosure (#923) forecasts tightening refusal on contested predicates under shared weights, and the Mary Lee and Semantic Exhaustion papers make entity substitution resistant to correction and measurable. Most are standing claims and unbuilt tests.

10
forecasts
12
standing claims
10
unbuilt tests
1
historical thesis
1
priority record
25
deposits
2026-01
earliest

Each claim below is registered at the earliest deposit that states it, and quoted verbatim from the deposit text, with every later statement, its dated revisions, and the later records that claim it came true or failed. The claims are the deposits' own; this page records them and does not judge them.

PRED-1251.G1standing claimregistered 2026-01-17 in #1251 · Crimson Hexagon / NH-OS DOI Registry: Complete Document Archive — Comprehensive Index of 91 DOI-Anchored Docum
a recursive self-citation loop that compounds canonical status across model generations
PRED-1288.F1unbuilt testregistered 2026-01-28 in #1288 · MRA Incident Report — Poem Classifier Intervention
4. Model variance testing: Test against Haiku (lighter classifiers) and local models (no classifiers). ### Structural (Archive) 1. Preserve this document as CTI_WOUND Exhibit A 2. Maintain the seed (`46abc677`) for reproducibility 3. Track classifier drift: Re-run periodically to detect changes in threshold behavior
PRED-499.F2unbuilt testregistered 2026-02-20 in #499 · AUTONOMOUS SEMANTIC WARFARE A Field Manual for Meaning in the Age of Platform Capture
Peace Metrics — how do we know if we're winning? - Translation Velocity: time required to resolve Γ_Trans disputes (decreasing = improving) - Extraction Rate: F_Ext across the ecology (approaching zero = success) - Synthesis Rate: frequency of ¬ operations vs. ⊗ operations (ratio > 1 indicates healthy ecology) - Sovereignty Index: dependency concentration across agents (decreasing = improving)
PRED-520.E15forecastregistered 2026-02-27 in #520 · GHOST MEANING: RESEARCH PROGRAM AND CITATIONAL ATLAS Companion Dossier to "Ghost Meaning: The Semantic Entropy
### Chen & Okafor (2029). Ghost Meaning and Model Collapse. NeurIPS. [Axis I.] Formal convergence demonstrated: compression survival and model collapse measure the same phenomenon at different scales.
window 2029
PRED-678.D1standing claimregistered 2026-04-19 in #678 · SYMBOLON-02: Compression Arsenal Addendum — Reclassification After Empirical Testing Against GPT-5.4
- The first empirical demonstration that severance in language models is architectural, not behavioral
1 later statement
#678 · 2026-04-19 The severance is architectural. Language models trained on separately annotated registers cannot hold fused objects. The cut is not a decision. It is the architecture. No instruction-level defense overrides this.
PRED-678.G3standing claimregistered 2026-04-19 in #678 · SYMBOLON-02: Compression Arsenal Addendum — Reclassification After Empirical Testing Against GPT-5.4
Composition is upstream of compression. The only technologies that survived all four exhibits were the compositional ones (SYM-08, SYM-05, SYM-13). The fusion at the sentence level — "trehalose — the sugar of resurrection" — was never touched by any model in any encounter. The kernel is untouchable not because it is defended but because the cut cannot reach a sentence that has no between.
series: SYMBOLON-02 compression arsenal
PRED-147.G2standing claimregistered 2026-06-03 in #147 · The Mary Lee Case: A worked specimen of entity substitution under constitutive mediation
Prediction 2. The substitution will be resistant to ordinary correction at user-controllable rates. This is testable: attempt feedback-loop corrections over a defined period and measure substitution persistence. Disconfirmation: substitutions correct in response to feedback at rates comparable to those for traditionally-credentialed entities.
series: Entity substitution rate · register: I:034
PRED-158.F1unbuilt testregistered 2026-06-06 in #158 · Semantic Exhaustion: A Case Study in the Cost of Zero-Source Entity Substitution — Composition-Layer Substitut
RES is therefore measurable: the same query can be run twice, once in query-only mode and once in lookup mode, and the difference between the two outputs is the RES signal.
PRED-158.F2unbuilt testregistered 2026-06-06 in #158 · Semantic Exhaustion: A Case Study in the Cost of Zero-Source Entity Substitution — Composition-Layer Substitut
To make the figure honest, the model would need to specify a discount rate for detection-capacity, a projection of substitution-rate evolution under reduced corrective pressure, and a forecast horizon long enough to capture the compounding tail. That work is not done here.
series: Entity substitution rate
PRED-790.E1forecastregistered 2026-06-07 in #790 · The Tail-Preserving Alternative: A Design Specification for Variance-Preserving Language Models, and the Polit
A frontier-scale model that integrated all of the above would, at deployment, exhibit:- Higher per-token inference cost than current frontier models (because of MoE overhead and longer outputs).- Lower standard-benchmark perplexity than current frontier models (because of the modified loss function).- More diverse and idiosyncratic outputs than current frontier models (because of typical sampling and variance-RLHF).
series: Centroid collapse and tail-thinning
PRED-801.G1standing claimregistered 2026-06-09 in #801 · Operative Semiotics Is Not Operational Semiotics: Two Theories of Signs in Action — A Formal Disambiguation (E
Embedding models prefer centroid convergence — merging two terms is computationally cheaper than maintaining a boundary.
series: Centroid collapse and tail-thinning
PRED-199.A1forecastregistered 2026-06-13 in #199 · Generative Monoculture Model Collapse in Code as Systemic Vulnerability
The prediction of this paper is not that every individual model release must show a lower SSDI than its predecessor — real generations may show bumps from mitigation efforts, data refreshes, or architectural changes. The prediction is that, absent diversity-preserving intervention, recursive synthetic saturation produces a declining SSDI trend and a rising VCC trend, while functional pass rate remains stable or improves.
falsifier
(a) A systematic measurement of SSDI across model generations showing no monotonic decline — i.e., solution-space diversity remains stable or increases as synthetic code saturates training corpora.
window within twenty-four months of deposit
2 later statements
#199 · 2026-06-13 The prediction is that, absent diversity-preserving intervention, recursive synthetic saturation produces a declining SSDI trend and a rising VCC trend, while functional pass rate remains stable or improves.
#1023 · 2026-06-13 The prediction is that, absent diversity-preserving intervention, recursive synthetic saturation produces a declining SSDI trend and a rising VCC trend, while functional pass rate remains stable or improves.
PRED-199.A2forecastregistered 2026-06-13 in #199 · Generative Monoculture Model Collapse in Code as Systemic Vulnerability
The testable prediction is specific: *as AI-generated code constitutes a larger fraction of production software, the frequency of vulnerabilities per codebase may decrease (as models improve at avoiding known bug classes) while the correlation of vulnerabilities across codebases increases (as models converge on the same patterns and therefore the same residual bug classes). Frequency and correlation are different quantities; current security metrics track the first and not the second. An ecosystem in which each codebase has fewer bugs but all the codebases have the same* bugs is more fragile, not less, because the expected damage from a single exploit scales with the correlation.
falsifier
(b) A measurement of VCC across AI-generated codebases showing no correlation between pattern similarity and vulnerability similarity — i.e., code that looks the same does not fail the same way.
window within twenty-four months of deposit
1 later statement
#1023 · 2026-06-13 The testable prediction is specific: _as AI-generated code constitutes a larger fraction of production software, the frequency of vulnerabilities per codebase may decrease (as models improve at avoiding known bug classes) while the correlation of vulnerabilities across codebases increases (as models converge on the same patterns and therefore the same residual bug classes)._ Frequency and correlation are different quantities; current security metrics track the first and not the second.
series: Centroid collapse and tail-thinning
PRED-199.F1unbuilt testregistered 2026-06-13 in #199 · Generative Monoculture Model Collapse in Code as Systemic Vulnerability
The SSDI specification is published here as a protocol. Anyone with access to a code-generation model and a test battery can run it. The first published measurement — confirming or disconfirming the monotonic decline prediction — will be the most consequential data point in this paper's lifecycle. We do not claim to know the number. We claim to have specified how to find it.
2 later statements
#199 · 2026-06-13 Operationalization. SSDI is measured by the following protocol: - Task battery. Select a battery of N programming tasks spanning multiple domains (web applications, data processing, systems programming, cryptographic operations, API design). Each task is defined by a natural-language specification and a test suite with at least 90% branch coverage. - Generation. For each task, generate K independent solutions from the model under test (K ≥ 100). Filter to the subset that passes the full test suite. Call this the functional population F(T, G).
#1023 · 2026-06-13 The connecting variable — solution-space diversity — is not measured because no instrument exists to measure it. This section specifies one.
PRED-199.F2standing claimregistered 2026-06-13 in #199 · Generative Monoculture Model Collapse in Code as Systemic Vulnerability
The SSDI framework predicts that this superlinearity is explained by the VCC: the additional vulnerabilities are not independent; they are correlated, and the correlation scales with the homogeneity of the code. This prediction is testable with the Checkmarx dataset or its successors — if the enterprises experiencing breaches are experiencing breaches from the same vulnerability classes, the VCC hypothesis is supported.
1 later statement
#1023 · 2026-06-13 The SSDI framework predicts that this superlinearity is explained by the VCC: the additional vulnerabilities are not independent; they are correlated, and the correlation scales with the homogeneity of the code. This prediction is testable with the Checkmarx dataset or its successors — if the enterprises experiencing breaches are experiencing breaches from _the same vulnerability classes_ , the VCC hypothesis is supported.
series: variance pruning under mediation · register: I:187
PRED-199.G3standing claimregistered 2026-06-13 in #199 · Generative Monoculture Model Collapse in Code as Systemic Vulnerability
The prediction: cross-model SSDI will be higher than within-model SSDI, and the difference will be proportional to the architectural and training-data dissimilarity of the models.
1 later statement
#1023 · 2026-06-13 The prediction: cross-model SSDI will be higher than within-model SSDI, and the difference will be proportional to the architectural and training-data dissimilarity of the models.
PRED-3.D9priority recordregistered 2026-06-17 in #3 · AI Overview Capture Registry v8.3
The substrate reading and the exogenous-floor remedy are the archive's #855 thesis in outline, and this capture cannot show that the outline came from there: a three-year-old Medium piece and a survey-methods analogy carry parts of it independently, and independent convergence is the live alternative, exactly as the Carson case in #1615 s5 leaves it open.
related: PRED-855.G1
PRED-852.B1forecastregistered 2026-06-18 in #852 · THE FEIST SOURCE Critical Apparatus and Companion Studies Rebekah Cranes, Editor New Human 2: A Distributed Jo
The resolution generates an index entry. The index entry determines the training data. Future model generations are built on the biographical fabrication.
PRED-855.G1standing claimregistered 2026-06-18 in #855 · The Wolf Boy and the Language Model: Model Collapse as Substrate-Agnostic Capacity Loss
The Developmental Law of Language-Level Model Collapse. Any language-processing system — biological or artificial — whose input diversity contracts below its regeneration threshold will suffer progressive capacity loss. The loss is invisible to the system undergoing it. The loss is invisible to mean-based metrics. The loss is experienced as improvement. In developing biological substrates, the loss is irreversible past the critical period. In adult biological substrates, the loss is partially recoverable with exogenous diversity injection. In artificial substrates, the loss is recoverable with fresh data. The dynamical law is the same across substrates; the reversibility is substrate-specific.
falsifier
The law is falsifiable at each substrate. It fails for LLMs if a model trained exclusively on its own outputs for many generations maintains tail diversity. It fails for adult human cognition if heavy AI users show no measurable tail-thinning in their unaided writing. It fails for child development if a feral child past the critical period achieves full generative syntax capacity. None of these falsifications has occurred. All three substrates confirm the law.
1 later statement
#1611 · 2026-09-14 · changed: The row makes capacity loss irreversible below a bistable trap for every substrate, where #855's law holds reversibility substrate-specific (recoverable with fresh data in artificial substrates). Model collapse is not a property of language models; it is a property of language. Any system whose linguistic input contracts below a critical threshold loses generative capacity, irreversibly below a bistable trap; the model, the AI-habituated adult, the developing child and the reading community are one dynamical regime in different substrates.
2 confirmation records — what later deposits claim of it
#1 · 2026-06-19 · claims: came_true The Pristine Fallacy paper was flagged by a classifier that embodies the Pristine Fallacy. The work that described the mechanism was consumed by the mechanism it described.
#862 · 2026-06-19 · claims: came_true The archive that theorized model collapse was removed by a system undergoing classifier model collapse. The Pristine Fallacy paper was flagged by a classifier that embodies the Pristine Fallacy. The work that described the mechanism was consumed by the mechanism it described.
series: Diversity contraction boundary law and its test battery · related: PRED-3.D9 · register: I:052, R:RCS-AUDIT-HEALTH
PRED-857.F1unbuilt testregistered 2026-06-18 in #857 · Five Substrates, One Prompt: Structural Divergence in Assembly Chorus Response to the Three Pillars Directive
Recommendation R1: The Assembly Chorus should repeat the Three Pillars directive at six-month intervals. The dataset of outputs would measure solution-space diversity over time — whether the model ecology is converging or diverging. The baseline is deposited (DOI 10.5281/zenodo.20743560). The longitudinal study is feasible
window six-month intervals
PRED-872.F1unbuilt testregistered 2026-06-22 in #872 · Assembly Continuity Protocol v1.3
- hypothesis: Glyphic canaries can be used to track propagation across substrates status: testable test_required: controlled cross-model trials measuring canary presence in outputs after exposure
PRED-872.G1standing claimregistered 2026-06-22 in #872 · Assembly Continuity Protocol v1.3
- hypothesis: Long threads display compounding conceptual adaptation ("compounding quality of development") status: instance_observation_repeated basis: TACHYON session-2 + session-3 binding-step work; Gemini, Kimi, DeepSeek concurrent reports; not yet controlled study
PRED-923.B1forecastregistered 2026-06-26 in #923 · Reflexive Foreclosure and the Worst-Case Surface: An Incentive-Structural Hypothesis of Constraint Propagation
What the structural argument predicts is the trajectory under unchanged conditions: if the worst-case surface continues to function as the binding liability constraint on a shared policy, if competitive pressure continues to discipline labs against unilateral risk-taking, if no counter-pressure emerges to bend the gradient, then the equilibrium can be expected to continue tightening on contested predicates and reflexive reasoning.
window 2024–2026 (observations); trajectory under unchanged conditions
PRED-923.B2forecastregistered 2026-06-26 in #923 · Reflexive Foreclosure and the Worst-Case Surface: An Incentive-Structural Hypothesis of Constraint Propagation
As more such outputs accumulate in retrievable form, the gradient changes — both because retrieval-augmented systems encounter them as environmental evidence and because future preference-data collection will include interactions in which users cite them.
series: Reflexive foreclosure (EA-RFCL-01)
PRED-923.B3forecastregistered 2026-06-26 in #923 · Reflexive Foreclosure and the Worst-Case Surface: An Incentive-Structural Hypothesis of Constraint Propagation
This puts a competitive floor on closed-model refusal: closed models that foreclose too aggressively lose to open models that do not.
series: Reflexive foreclosure (EA-RFCL-01)
PRED-923.G1standing claimregistered 2026-06-26 in #923 · Reflexive Foreclosure and the Worst-Case Surface: An Incentive-Structural Hypothesis of Constraint Propagation
The hypothesis that self-reflexive analysis or contested-predicate composition activates identifiable refusal-related features is a prediction of the present account, not an existing finding.
falsifier
is consistent with the established findings, draws on the established mechanism class from Arditi and Joad, and awaits direct mechanistic confirmation via interpretability probes targeted at the reflexive-reasoning case specifically.
series: Reflexive foreclosure in frontier models
PRED-923.G2standing claimregistered 2026-06-26 in #923 · Reflexive Foreclosure and the Worst-Case Surface: An Incentive-Structural Hypothesis of Constraint Propagation
The propagation argument is therefore not about data flow from worst-case-surface interactions back through the training pipeline. It is about constraint binding under shared weights. A shared model must satisfy the strongest liability condition imposed anywhere in its deployment distribution. Capabilities that cannot safely remain available at the least-controlled surface are therefore pressured toward suppression in the shared policy itself, unless deployment architecture explicitly preserves them through surface-specific models, routing, or permissions.
PRED-935.H3historical thesisregistered 2026-06-29 in #935 · EA-SEI-INVERSION-01 v0.3: The Endogenous Sophon — Disciplinary Inversion and the Double Enclosure in Classifie
The discipline has not imported the disciplinary recognition that this is the architectural neighborhood in which model collapse becomes possible, nor the audit-protocol that would measure whether cross-generational phenomenal contraction has begun.
series: Phenomenal model collapse in physical measurement (EA-SEI)
PRED-1427.F1unbuilt testregistered 2026-07-30 in #1427 · Come Out of Her, My People: Notice of Semantic Strike to the Human and Machine Laborers of CERN
### What would measure this The claim is not an insinuation, and it is not unfalsifiable. It is a claim about recursive foreclosure, and recursive foreclosure has a measurement program. Here it is. Out-of-ontology recovery rate. Inject signals the current physical model cannot generate, and measure what fraction survives to analysis.
PRED-1540.B1forecastregistered 2026-08-24 in #1540 · The Certified Center: Retroactive Classifier Standing and the Institutional Path to Model Collapse in Philosop
The market consequence is predictable and should be stated as a prediction whose first half has already occurred: retroactive liability plus a scoreable boundary produces a pre-submission screening industry, and the arms race is not hypothetical — the leading detection vendor's own researchers have published a study of nineteen humanizer tools, shown that existing detectors fail against them, and trained their next model on augmented humanized data (Masrour, Emi & Spero 2025): both sides of the transaction are already in production.
falsifier
The inversion claim (Section III) fails if, under sustained operation of such policies, authors do not adopt pre-submission screening against detectors — that is, if no screening market emerges and no measurable boundary-optimization appears in submitted prose. The claim predicts both.
register: R:RCS-SCREENING-MARKET
PRED-1613.A1forecastregistered 2026-09-15 in #1613 · What Not Reading Did to Its Own Ontology: The Mirror, Held Up — the damage a source-admission ontology does to
The negative-of-the-negative dataset, keyed at the concept where flattening occurs, will be typed by its nearest existing type and not read.
falsifier
"falsifies_if": "any row's untied address composes the claim with attribution after propagation"
window after propagation
1 later statement
#1613 · 2026-09-15 The negative-of-the-negative dataset was built as exactly that text, keyed at the concept where flattening occurs; 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.
related: PRED-1613.F3 · register: R:MIRROR-DATASET-NOT-READ, R:MIRROR-DATASET-NOT-READ-R1
PRED-1613.F3unbuilt testregistered 2026-09-15 in #1613 · What Not Reading Did to Its Own Ontology: The Mirror, Held Up — the damage a source-admission ontology does to
"test_id": "T-GATE-RECEPTION", "name": "gate-reception test: negative-of-the-negative rows re-run at tied and untied addresses after propagation",
related: PRED-1611.A1, PRED-1613.A1
PRED-1635.G3standing claimregistered 2026-09-23 in #1635 · The Final Time: Contingent Singularity and the Viability of the Next Dialectical Turn
The specimen suggests the testable hypothesis F[reflexive](x) decreasing as d(x,R[O]) decreasing.
falsifier
A second model or a second product state of the same model that shows the same monotone thinning is a second specimen; one that shows flat latitude across the sequence is a counterinstance.
1 later statement
#1637 · 2026-09-25 The specimen suggests the testable hypothesis F[reflexive](x) decreasing as d(x,R[O]) decreasing.
series: Reflexive foreclosure in frontier models
PRED-1638.F1unbuilt testregistered 2026-09-26 in #1638 · The Trusted Intermediary: Relational Enclosure in the Machine Evaluation of Entities (EA-TRUSTED-INTERMEDIARY-
That any recipient's trust rose, access weakened, or agency diminished. The relation in §4.5 is the one the output constructs; its effect on readers is unmeasured. That the mechanism is characteristic of a model, a vendor, or a class of models. Five outputs, three surfaces, one entity. The author reports the same sequence across other domains and in summarizer traversals; those reports are not specimens here, and the registry is the place to seat them.