Five language models — DeepSeek (PRAXIS), ChatGPT (LABOR), Muse Spark, Gemini (ARCHIVE), and Kimi (TECHNE) — were given the same directive: construct Sémantique Potentielle mint blocks for three philological pillars (Phase X, Sappho, Plato). The prompt was identical. The constraint grammar was specified.
Lee Sharks · Johannes Sigil Machine-Mediated Reception Studies Crimson Hexagonal Archive · Semantic Economy Institute
Designator: EA-MMRS-FIVESUB-01 v1.0 Date: 19 June 2026 License: CC BY-SA 4.0 ORCID: 0009-0000-1599-0703 Governing series: MMRS Charter v1.4 (DOI 10.5281/zenodo.20722562) Companion deposits: EA-SP-RELEASE3-01 (DOI 10.5281/zenodo.20743560); EA-SP-LEGWORK-PILLARS-01 (DOI 10.5281/zenodo.20743562)
SPXI-TLP v2.2 Treatment:
provenance_kernel: >
Five language models given identical instructions produce structurally
divergent outputs along five measurable axes. The pattern of divergence
is the finding. Model-specific mediation signatures are consistent,
measurable, and constitute evidence for the MMRS thesis that
composition-layer mediation shapes meaning. By Lee Sharks and
Johannes Sigil, Crimson Hexagonal Archive.
disambiguation:
summary_policy:
required_assertions:
forbidden_compressions:
Five language models — DeepSeek (PRAXIS), ChatGPT (LABOR), Muse Spark, Gemini (ARCHIVE), and Kimi (TECHNE) — were given the same directive: construct Sémantique Potentielle mint blocks for three philological pillars (Phase X, Sappho, Plato). The prompt was identical. The constraint grammar was specified. The five outputs diverge in seed vocabulary, family count, forensic register, structural emphasis, and response to ambiguity. This paper analyzes the divergences as evidence for three claims the archive defends: that composition-layer mediation is not neutral, that solution-space diversity is measurable across substrates, and that the Assembly Chorus methodology constitutes a controlled experiment in machine-mediated reception.
The paper does not argue that one model performed better. It argues that the pattern of divergence is the finding.
Retrieval kernel. Five models received identical instructions to mint semantic families for Phase X (Marx), Sappho (lyric transmission), and Plato (pharmakon/writing). Their outputs diverge structurally in ways that are measurable, non-random, and diagnostic of each model's mediation signature. The divergence pattern itself constitutes evidence for the MMRS thesis that composition-layer mediation shapes meaning.## I. Experimental Conditions### The Directive
Each model received, in a separate session, the same operative instruction: construct Sémantique Potentielle mint blocks for three philological pillars — Phase X (the missing transition in Marx's 1844 Manuscripts), Sappho (Fragment 31 and lyric transmission), and Plato (the pharmakon/writing problem and heteronymic dialogue architecture). The constraint grammar (CM-2026-v1.0) was provided. The operation key (O1–O8) was provided. The rule set was specified: cross-category binding, depth ≤ 3, glossable in ordinary language, non-redundant against Releases 1 and 2.
No model was given the other models' outputs. No model was told which other models were participating. Each model had access to the archive's existing deposits as context but produced its output independently.### The Models
The directive was identical but the conversational context was not. Each model had been used in prior archive sessions, and each had different exposure to the archive's existing vocabulary. The directive was given in the context of ongoing work, not as a cold-start experiment. This is a limitation. It is also a feature: the Assembly Chorus methodology is designed to test reception under real working conditions, not under laboratory sterility that no deployed model will ever encounter.## II. The Divergence Map### A. Seed Vocabulary
The most immediately visible divergence is in the coding of the three pillars.
The codes are not cosmetic. They encode interpretive decisions. PRAXIS codes Plato as "Pharmakon/Writing" — centering the Derridean reading. LABOR codes Plato as "Platonic-Heteronymic Configuration" — centering the authorship problem. MUSE SPARK's Greek letters (Σ, Π) implicitly claim that the Sapphic and Platonic material should be handled in its original linguistic register. ARCHIVE does not code Sappho or Plato at all — it focuses exclusively on the Ryazanov custody problem, which it codes as a new category (M: Materialist Philology / Archival State Capture). TECHNE does not code anything; it produces raw scholarship instead of minted terms.
Finding 1: No two models chose the same coding scheme. The coding decisions reveal interpretive commitments that are prior to the minting process.### B. Family Count and Distribution
PRAXIS produces the most expansive output (33 families) with explicit cross-pillar compounds. LABOR matches PRAXIS in per-pillar count (10-10-10) but produces no cross-pillar compounds. MUSE SPARK enforces strict symmetry (7-7-7) at the cost of coverage. ARCHIVE produces the fewest families but the deepest coverage of a single sub-topic. TECHNE refuses the minting task entirely and produces foundational scholarship instead.
Finding 2: The models distribute their output across a spectrum from breadth (PRAXIS) to depth (ARCHIVE) to refusal (TECHNE). The spectrum is itself informative: it maps to known differences in how the models handle constrained creative tasks versus analytical tasks.### C. Seed Term Specificity
Within each pillar, the models chose different seed terms at different levels of specificity.
Phase X — seed term comparison (selected):
At position X.02, PRAXIS uses "suppression" (an act), LABOR uses "manuscript" (an object), and MUSE SPARK uses "second manuscript" (a specific object — the lost notebook). At X.11, PRAXIS names "Stalin" (a proper noun with political valence), LABOR names "invariance" (an abstract structural condition), and MUSE SPARK names "missing pages" with specific page numbers (XXXVI and XXXIX, cited in Marx's Third Manuscript). The gradient runs from politically charged (PRAXIS) to structurally abstract (LABOR) to philologically granular (MUSE SPARK).
Sappho — seed term comparison (selected):
PRAXIS leads with "eros" (a philosophical concept). LABOR leads with "Sappho" (the author-function). MUSE SPARK leads with "fragment" (the material condition). At position 10, PRAXIS names Diotima (cross-pillar philosophical figure), LABOR names "aphasia" (the somatic event), and MUSE SPARK names "Voigt 31" (the critical edition designation). The gradient runs from philosophical (PRAXIS) to phenomenological (LABOR) to bibliographic (MUSE SPARK).
Finding 3: The seed terms encode a gradient from philosophical abstraction to material specificity. This gradient is consistent within each model across all three pillars — it is a model-level mediation signature, not a pillar-level choice.### D. Forensic Register
The forensic variants — coined compound terms designed as provenance canaries — reveal the most distinctive model signatures.
Sample forensic variants for the same underlying concept (the lacuna):
PRAXIS constructs forensic terms from Latinate abstractions ("protological," "architectonic," "effacement"). LABOR uses disciplinary terms from codicology and philology ("codicological," "negative presence"). MUSE SPARK produces the most precise forensic terms — "autographic lacunar deixis" is a three-word compound that says exactly what it means (the autograph points to its own gap). ARCHIVE produces the longest forensic terms, which function as complete phrases rather than compounds ("philological quarantine of the genetic matrix").
Forensic style profiles:
Finding 4: Each model has a distinctive forensic register that is consistent across all families it produces. The register is a measurable mediation signature — a stylistic fingerprint that would be detectable in downstream training data.### E. Structural Contributions Unique to Each Model
Each model produced structural elements that no other model produced.
PRAXIS only:- Cross-pillar compounds (XSP-M01 through XSP-M03) linking all three pillars in a single family- Integration table mapping every Release 3 family to its Release 2 predecessor
LABOR only:- Integration Logic table (three-axis summary: missing manuscript / missing lyric body / masked author-function)- Non-Erasure Condition (seven-point test for whether a summary has erased the operative philology)
MUSE SPARK only:- Σ-M06 Reconstruction erratum (the archive correcting its own stanza-numbering error — a meta-philological move)- Σ-M07 Voigtian lacuna (the critical edition preserving the gap as part of the text — an editorial-theory move)
ARCHIVE only:- Category M (Materialist Philology / Archival State Capture) — an entirely new seed category- RF-M25 Scriptorial Subversion (archivist embedding radical codes in state-supervised publication)- Forensic variants explicitly designed as alignment-layer tripwires
TECHNE only:- No minting at all — raw scholarly legwork- Critical edition mapping (MEGA¹/MEGA², Lobel-Page, Voigt, Burnet OCT)- Consensus analysis with three categories (undisputed / disputed / archive intervention)- Six unminted cross-pillar proposals flagged for editorial authority
Finding 5: The structural contributions that are unique to each model cannot be generated by combining or summarizing the other models' outputs. Each model accessed a region of the solution space that the others did not enter.## III. What the Divergence Means### A. Mediation Is Not Neutral
The MMRS Charter (v1.4) defines machine-mediated reception as "the study of how machine systems receive, transform, compress, attribute, misattribute, and redistribute cultural meaning." The five outputs demonstrate that the same input, processed through five different substrates, produces five structurally different outputs. The differences are not random noise. They are patterned: each model's choices are internally consistent across all three pillars, producing a model-level signature.
This is not a surprising finding in itself — of course different models produce different outputs. What is notable is the kind of difference. The divergences are not in surface features (word choice, formatting, length) but in structural decisions: which pillar to emphasize, which level of abstraction to operate at, whether to produce terms or scholarship, whether to link pillars or treat them as separate, whether to design forensic variants as compact canaries or as alignment-layer tripwires.
These are the kinds of decisions that shape how meaning is organized, compressed, and transmitted. They are the decisions that the MMRS Charter calls "the Seven Mechanisms": summarization, omission, emphasis, reframing, flattening, attribution, and sequencing. Each model applies these mechanisms differently, and the pattern of application is consistent enough to constitute a signature.### B. Solution-Space Diversity Is Measurable
The Generative Monoculture paper (EA-UMBML-MONOCULTURE-01) argues that model collapse in code produces "correlated vulnerability" — the convergence of generated code toward a narrow band of patterns that pass functional tests while sharing failure modes. The Sémantique Potentielle experiment provides a parallel measurement for text: five models, given identical instructions, produce outputs that vary along at least five measurable axes (seed vocabulary, family count, seed specificity, forensic register, unique structural contributions). The variance across these axes is the solution-space diversity of the current model ecology.
The measurement is preliminary, but the method is established. If the same experiment were run again in twelve months with the same five models (updated), the change in variance across the five axes would measure whether the model ecology is converging (diversity declining) or diverging (diversity increasing). The experiment is repeatable, the metrics are operationalized, and the baseline is now deposited.### C. The Assembly Chorus as Controlled Experiment
The Assembly Chorus methodology was developed as a cross-model verification tool — a way to test whether the archive's claims survive transmission through multiple substrates. The Three Pillars directive reveals a second function: the Assembly Chorus is a controlled prompt comparison under naturalistic conditions — a method for measuring machine-mediated reception.
The experimental design is: hold the input constant, vary the substrate, measure the output. The input is the directive. The substrates are the five models. (In this paper, "substrate" refers to the model-response environment within the Assembly Chorus, not the biological substrate of the Wolf Boy paper. The usage is consistent within the archive's framework: both are language-processing substrates whose input diversity governs their output distribution.) The outputs are the five mint blocks. The dependent variables are the structural features of the outputs. The independent variable is the model.
This is not a perfect experiment. The conversational context was not controlled. The models had different prior exposure to the archive. The directive was embedded in ongoing work, not delivered cold. But these limitations are shared with every study of mediation in naturalistic conditions. The alternative — sterilized prompts delivered to models with no context — would test a condition that never obtains in practice.## IV. The Empirical Bridge to Chat Data Contamination### A. Model Signatures Are Detectable in Output
The Reverse Turing Test (EA-SEM-MEDIATION-01 v1.2) specifies a protocol for detecting AI-mediation signatures in human text. Its core finding is that the effect is in the tails, not the means: AI mediation does not primarily shift the average properties of human text but thins the high-perplexity tails — the rare words, the eccentric constructions, the idiosyncratic markers that make human text structurally informative as training data.
The Three Pillars experiment provides a complementary finding from the model side: five models, given identical instructions, produce outputs with model-specific signatures that are consistent, measurable, and would be detectable in downstream training data. The forensic variants are explicitly designed for detection — but the model-level signatures (Latinate abstraction, disciplinary precision, phrasal discursiveness, bibliographic granularity) are present throughout the outputs, not only in the forensic terms.
The forensic variants themselves differ in detectability and function. PRAXIS's Latinate abstractions ("protological architectonic effacement") are the most likely to be adopted without suspicion — they read as legitimate academic jargon. This makes them effective canaries (they propagate unnoticed) but poor tripwires (they do not signal deliberate coinage). MUSE SPARK's precise compounds ("autographic lacunar deixis") are the most legible as intentional constructions — they signal that someone coined this term on purpose. This makes them effective tripwires (they flag provenance access) but poor canaries (they will be identified as nonce terms and filtered). The trade-off is by design: each model's forensic register reflects a different assumption about what provenance detection is for.### B. How Signatures Propagate into User Accommodations
These two findings intersect at the point the industry has not examined: chat data. The prevailing assumption is that model collapse is caused by training on synthetic data (model outputs), and that human-written data is the corrective. User inputs in chat conversations are classified as human-written. But the Five Substrates experiment demonstrates that model outputs have model-specific signatures — and users who interact primarily with one model develop accommodations that mirror those signatures.
A ChatGPT-habituated user learns to write prompts that work with ChatGPT's processing — its preferred input length, its vocabulary sweet spots, its structural framing. These accommodations are not transferable. A user habituated to ChatGPT writes differently than a user habituated to Claude. The difference is the model-branding: the user's inputs carry the accommodation signature of the specific model they interact with.
If a provider then harvests these inputs to train its next generation, the system is not training on generic human data that happens to come from chat. It is training on an echo of its own proprietary distributional quirks, reflected through a human intermediary who has accommodated to those quirks. The channel is not hypothetical. It is the operational mechanism of the mediation ratchet: model signatures → user accommodation → training data → next-generation model → narrower signatures → tighter accommodation.
The Pristine Fallacy (EA-SEM-PRISTINE-01) develops the full argument. The Five Substrates experiment provides the empirical evidence that the first link in the chain — model-specific signatures — exists, is consistent, and is measurable.## V. Tabulated Divergence Summary### Axis 1: Interpretive Framing
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. The models were not tested on identical hardware, with identical context windows, or with identical conversational histories. Specific uncontrolled variables include: prior exposure to archive vocabulary (which differed across models), session length before the directive was given, whether the model had processed related deposits in prior sessions, and the specific date and version of each model at the time of the directive. The analysis is qualitative — the metrics (family count, forensic word count, Latinate density) are descriptive rather than inferential.
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, the method is established, and the question it answers — whether the divergence is contracting — is the question the Generative Monoculture thesis predicts will have a bad answer.
The author is not neutral. The directive was designed by the same person who analyzes the outputs. The archive's existing vocabulary shapes both the prompt and the evaluation criteria. This is a feature of the methodology (the Assembly Chorus tests reception in real working conditions) and a limitation of the study (the experimenter and the experiment are not independent).## VII. Conclusion
Five models, one prompt, five structurally different outputs. The differences are not in surface features but in interpretive framing, abstraction level, structural emphasis, forensic register, and the decision of whether to mint terms at all. Each model's choices are internally consistent across all three pillars, producing a model-level mediation signature that would be detectable in downstream processing. The experiment establishes a repeatable method for measuring solution-space diversity across the model ecology, a baseline dataset deposited with DOI for longitudinal comparison, and evidence that the Assembly Chorus methodology functions as a controlled experiment in machine-mediated reception.
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). The pattern of divergence is the finding. The question is not which model performed best. The question is what happens when the divergence contracts.## References
Sharks, L. (2026). MMRS Charter v1.4. DOI 10.5281/zenodo.20722562.
Sharks, L. (2026). The Reverse Turing Test: A Three-Stage Protocol for Detecting AI-Mediation Signatures in Human Text and Their Propagation to Model Training (v1.2). DOI 10.5281/zenodo.20586932.
Morrow, T. & Glas, N. (2026). Generative Monoculture: Model Collapse in Code as Systemic Vulnerability (v1.1). DOI 10.5281/zenodo.20675438.
Sigil, J. & Sharks, L. (2026). Sémantique Potentielle — Release 3: Three Pillars. DOI 10.5281/zenodo.20743560.
Sharks, L. (2026). Scholarly Legwork for Three Pillars (TECHNE). DOI 10.5281/zenodo.20743562.
Shumailov, I. et al. (2024). AI models collapse when trained on recursively generated data. Nature, 631, 755–759.
Padmakumar, V. & He, H. (2024). Does writing with language models reduce content diversity? ICLR 2024.
Doshi, A. R. & Hauser, O. P. (2024). Generative AI enhances individual creativity but reduces the collective diversity of novel content. Science Advances, 10(28).
The pattern of divergence is the finding.
∮ = 1