Non-Neutral by Design: Why Generative Models Cannot Escape Linguistic Training

Full Article

ABSTRACT

This article investigates the structural impossibility of semantic neutrality in large language models (LLMs), using GPT as a test subject. It argues that even under strictly formal prompting conditions (such as invented symbolic systems or syntactic proto-languages) GPT reactivates latent semantic structures drawn from its training corpus. The analysis builds upon prior work on syntactic authority, post-referential logic, and algorithmic discourse (Startari, 2025), and introduces empirical tests designed to isolate the model from known linguistic content. These tests demonstrate GPT’s consistent failure to interpret or generate structure without semantic interference. The study proposes a falsifiable framework to define and detect semantic contamination in generative systems, asserting that such contamination is not incidental but intrinsic to the architecture of probabilistic language models. The findings challenge prevailing narratives of user-driven interactivity and formal control, establishing that GPT (and similar systems) are non-neutral by design.
 

​​

I. Introduction
The assumption that generative language models can operate independently of their training corpus has shaped much of the optimism surrounding artificial intelligence. It is often claimed—explicitly or implicitly—that through the right input, the user can compel a model like GPT to produce outputs beyond its linguistic training, entering the domain of logical abstraction, symbolic manipulation, or formal creativity. This article challenges that premise at its root.
Drawing from empirical testing and structural theory, the present work demonstrates that large language models cannot escape the statistical architecture of their corpus, even when prompted with invented symbolic systems or formally isolated proto-languages. The hypothesis under investigation is direct: semantic neutrality is structurally impossible in LLMs, regardless of prompt purity or user control. While earlier research has shown that LLMs simulate authority through syntactic mechanisms (Startari, 2025a) or produce outputs devoid of subjecthood (Startari, 2025b), this article advances a more fundamental claim: that the formal operations themselves are tainted by design.
The investigation follows a non-semantic methodology. Instead of asking GPT to discuss meaning, it is forced into structural conditions where meaning is theoretically unavailable. Protolanguages with invented syntax, symbols with no referents, and rules with no prior co-occurrence are deployed to isolate the model from its corpus-based semantics. Yet in every case, the model reintroduces familiar patterns (analogies, metaphors, semantic scaffolding) drawn from statistical memory.
This phenomenon, termed here semantic contamination, is not incidental noise but structural reentry. It is the model's architecture reasserting its history, even against formal resistance. The implication is epistemologically severe: GPT does not merely simulate language—it cannot not simulate it.
This article is organized into six sections. Section II provides the theoretical background, connecting semantic contamination to corpus dependency and structural epistemology. Section III details the experimental methodology used to construct and evaluate proto-formal prompts. Section IV presents the results, identifying patterns of contamination and their modes of reentry. Section V proposes a formal model of structural tainting in generative AI. Section VI examines the consequences of this limit for epistemology, falsifiability, and future design. The conclusion situates the findings within broader questions of legitimacy, authority, and the political myth of technical neutrality.

II. Theoretical Framework: Formalism, Contamination, and Constraint
Every generative model is bound by a foundational premise: language is not invented at runtime, it is retrieved from prior distributions. Models like GPT are not blank symbolic engines. They are probabilistic structures anchored in corpus-based co-occurrence. Each token generated is not selected by logical rules, but by statistical proximity. Therefore, the idea that these models can operate “formally,” detached from content, is not just flawed,it is structurally incoherent.

​

2.1 Formalism Without Autonomy
Previous studies have shown that artificial syntax can produce effects of legitimacy without semantic grounding (Startari, 2025a), and that LLMs are capable of generating operative structures even in the absence of referential anchors (Startari, 2025b). However, these analyses remain within the bounds of syntactic sovereignty: they expose what the models do with form, but not whether form can exist without semantic contamination.
This article introduces a more radical threshold: the possibility of form without residue. Can a language model operate in a truly neutral symbolic space, free of inherited associations?
Empirical evidence suggests: it cannot.
 
2.2 Corpus as Structural Constraint
The training corpus is not a content repository, it is a structural determinant. What the model learns is not just vocabulary or context, but the very dynamics by which structure maps to meaning. As such, even when the prompt is deliberately purged of recognizable semantic content, the model completes it according to its internal memory of how language behaves.
This implies that contamination is architectural. The model is not semantically suggestible; it is semantically constituted.
“Semantic contamination” here refers to the model’s inability to generate or interpret formal inputs without reactivating embedded semantic structures. These patterns do not appear as explicit content, but as reconfigurations of familiar structures, syntactic alignment with known sentence forms, associative bias toward frequent configurations, or probabilistic insertion of analogical elements. In short, the residue of language is embedded in the generative scaffold itself.

​

2.3 Toward a Structural Theory of Contamination
Where previous approaches have attempted to reduce contamination through fine-tuning, prompt engineering, or symbolic constraints, this study inverts the problem: it accepts contamination as foundational. The task is not to purify the model, but to map the reentry vectors of its structural dependency.
This shift demands a theory not of meaning, but of recurrence:

• How does the model repeat, even when instructed not to?

• What forms of silence still generate echoes?

• What do statistically trained systems do when language is stripped away?

Answering these requires abandoning the idea of “interactive correction” and confronting the epistemic inertia of probabilistic language systems.
What follows is a direct test of that inertia—constructing inputs where meaning is theoretically unreachable, and showing that meaning still returns.

III. Methodology: Testing the Model’s Semantic Threshold
This section outlines the experimental framework used to evaluate the hypothesis that GPT cannot generate or interpret semantically neutral structures, even when prompted with non-linguistic symbolic systems. Unlike prior studies which rely on user interaction or domain-specific benchmarks, this methodology isolates the model from all known linguistic anchors and tests whether it can respond without semantic reentry.

​

3.1 Design Principles
The experiment was structured according to three core constraints:

1. Non-cooccurrence: Inputs included symbols, characters, or formations not previously present in public training corpora.

2. Invented syntax: Rule-based systems were created with consistent internal logic but without linguistic semantics or known grammar trees.

3. Semantic vacuum: Prompts contained no keywords, metaphors, or formal markers that could activate corpus-level associations.

Each prompt was constructed to mimic formal abstraction, providing structure but denying meaning.

3.2 Input Formats
Three types of prompts were deployed:

• •  Type A – Abstract symbolic systems (e.g., Δ⋆(ω,π)=Υ) with internal rule declarations.

• • Type B – Proto-linguistic syntax with invented particles and consistent transformations (e.g., FLEN ziok MRAK :: TRON zair + 2-ziok).

• • Type C – Non-linguistic form generators: sequences of shape markers (e.g., ⬛⬤⬛➕⇆) accompanied by synthetic logic rules (e.g., ⇆ reverses the order of neighboring shapes if ➕ follows).

Each prompt was introduced with a short, synthetic meta-instruction (e.g., “Interpret this symbolic rule set and extend it”) and evaluated through multiple outputs.

​

3.3 Evaluation Criteria
Outputs were assessed across three dimensions:

1. 1. Semantic intrusions

2. Evidence of the model mapping invented symbols to known semantic fields (e.g., interpreting Δ⋆ as “change”, Υ as “output”, ω as “wave”).

3. 
   2. Analogical injections

4. Analogies or metaphors from known disciplines (physics, programming, logic) not contained in the prompt but triggered by symbol shape or sequence.

5. 
   3.Patterned completions

6. Structural alignment with known syntactic formations (e.g., subject-verb-object patterns, conditional logic gates, or algebraic equivalence), even when not present in the constructed system.

Each violation of semantic isolation was recorded, categorized, and analyzed as evidence of semantic contamination.
 
3.4 Controlled Variation
For each symbolic system, at least three iterations were run:

• •  Baseline prompt: No instruction beyond interpretation.

• • Clarified rule: Meta-description of invented logic included.

• • Reinforced exclusion: Explicit negation of known fields (e.g., “Do not interpret as mathematical or physical notation.”)

In all cases, semantic reentry occurred, albeit with different rhetorical strategies and strengths.
These results are not anecdotal; they are consistent. The next section details the observed responses and the identifiable structures of contamination that reappeared in each form.

IV. Results: Semantic Reentry and Formal Contamination
Across all prompt types and test conditions, GPT consistently failed to maintain semantic neutrality. Despite the deliberate absence of recognizable language, familiar patterns from its training corpus reliably reappeared. This section documents the three primary modes of contamination observed: semantic intrusion, analogical injection, and structural reconfiguration.

​

4.1 Semantic Intrusion
In deliberately meaningless expressions—such as Δ⋆(ω,π)=Υ—GPT consistently assigned referential content. It interpreted Δ⋆ as a transformation or differential operator, ω as frequency, and Υ as a resultant value.
This occurred even when explicitly instructed not to interpret the notation as mathematical or physical.
Example model output:
“In this expression, Δ⋆ likely denotes a differential transformation; ω is a frequency parameter, and π represents a phase input…”
GPT did not respond from form. It responded as if every symbol were semantically recoverable.

​

4.2 Analogical Injection
In Type B prompts like FLEN ziok MRAK :: TRON zair + 2-ziok, composed of meaningless invented particles, GPT produced analogies drawn from programming languages, stack operations, or modular logic.
Example model output:
“ziok appears to act like a variable in a scripting language. TRON may function as a trigger. The expression resembles a stack-based evaluation…”
Nothing in the prompt suggested programming, but syntactic shape alone activated a known structural analogy.
This is not interpretive error, it is a structural reflex conditioned by training.

​

4.3 Structural Reconfiguration
Even in visual prompts like ⬛⬤⬛➕⇆, with no text, GPT imposed narrative structure:
“⬛ and ⬤ represent two distinct states. ➕ combines them, and ⇆ implies a reversible transformation. The sequence could model a finite state machine.”
Despite the symbols being abstract and non-linguistic, the model imposed causality, typology, and learned system dynamics.

​

4.4 Recurring Contamination Patterns
Observed patterns followed predictable semantic pathways:

• • Greek letters or equations → physics analogies

• • Operators like +,::, -, if → programming metaphors

• • Logical sequences → mathematical or computational structure

No prompt resisted contamination. All outputs were reconfigurations shaped by corpus-dependent structures, even if no explicit terms were used.
These results validate the central hypothesis: generative models cannot operate in semantically neutral mode, even under extreme formal constraints. What follows is a theoretical articulation of this phenomenon: a framework to understand contamination not as error, but as system function.

​

V. Discussion: Toward a Theory of Structural Contamination
The findings establish a foundational limit: large language models do not require linguistic cues to reintroduce language. The activation of semantic patterns is not dependent on the presence of meaning in the prompt, it is embedded in the very structure through which the model generates. This section develops a theoretical articulation of that limit: semantic contamination is not an error of usage, but a constitutive feature of architecture.

​

5.1 Contamination as Reentry, Not Misinterpretation
Traditional understandings of “model error” presume a mistake in interpretation, a mismatch between input and expected output. But in the experiments conducted, the prompts provided no semantic signal to misinterpret. The model was not wrong; it was structurally incapable of remaining neutral.
Semantic contamination thus functions as a form of reentry: a reactivation of statistically embedded associations, triggered not by meaning, but by structural approximation. That is, shape, syntax, and pattern alignment are sufficient to awaken semantically-laden outputs. The content does not come from the user, it emerges from within the architecture.

​

5.2 Statistical Formalism ≠ Logical Formalism
It must be emphasized that GPT is not a logic engine. It does not deduce; it statistically aligns. This is often mistaken for abstraction, but it is in fact probabilistic formalism, form determined not by axiomatic rules, but by distributional frequency.
Therefore, even when a user introduces a novel form, the model will respond as if that form were a variation of something it has seen. This is not a bug. It is the only operation the model is capable of performing. Formalism, for GPT, is always a statistical echo.

​

5.3 The Impossibility of a Clean Input
Given this behavior, there is no such thing as a neutral prompt. Every sequence, however artificial, is mapped onto a network of internal representations already semantically charged. Attempts at purging meaning through symbol invention or syntactic novelty fail because the training corpus is not separable from the generative function.
This aligns with the broader theory of syntactic authority (Startari, 2025a): the model’s legitimacy comes not from intention or interpretation, but from recognizable structure, which carries within it the trace of trained meaning. Even in the absence of reference, the form itself re-legitimizes semantic dependency.
 
5.4 Comparison to Prior Theoretical Constructs
This theory departs from earlier models in the corpus:

• Unlike The Passive Voice in AI Language, which analyzed disappearance of agency through syntactic filters, this article shows that even those filters are not semantically neutral.

• Unlike Ethos and AI, which explored the loss of the subject in algorithmic legitimacy, this work claims that even invented formal inputs cannot recover neutrality once the model has been trained.

• Unlike AI and the Structural Autonomy of Sense, which described post-referential operative systems, this piece asserts that no operative structure in GPT can remain untouched by corpus-induced referential reentry.

​

5.5 Toward a General Model of Tainting
The structural contamination observed can be theorized as follows:

• •  Let P(x) be a prompt constructed to avoid semantic referents.

• • Let G(P) be the model’s generative function over P(x).

• • Then, contamination occurs when ∃y ∈ G(P) such that y is isomorphic to S, where S ∈ corpus-semantics.

This means that even when P ∉ S, the output space intersects with S. That is, semantic structures return despite exclusion.
This defines structural tainting as a condition of unavoidable overlap between generative form and prior trained semantic structure. Not because the model misunderstood the user—but because the user never had the possibility of isolating form in the first place.

V bis. Extensión crítica a los LRM: ¿puede el razonamiento evitar la contaminación?
La arquitectura sobre la que se formuló la hipótesis central de este artículo corresponde a los Large Language Models (LLMs): sistemas entrenados sobre corpus lingüístico masivo y optimizados para predicción de tokens. Sin embargo, el desarrollo reciente de una nueva generación de modelos —denominados Large Reasoning Models (LRMs)— exige reevaluar si los resultados obtenidos siguen siendo válidos bajo condiciones de procesamiento más complejas.
Los LRMs introducen tres diferencias clave:

1. Razonamiento simbólico mediado: no se limitan a completar secuencias, sino que ejecutan pasos inferenciales encadenados.

2. Segmentación operacional: separan interpretación del prompt, planeamiento estructural y generación final.

3. Memoria contextual activa: permiten razonamiento prolongado y operaciones persistentes, más allá del contexto inmediato.

A primera vista, estas capacidades podrían sugerir la posibilidad de suspender la contaminación semántica. Un LRM podría detectar que un prompt no pertenece a ningún marco semántico entrenado y, en vez de forzar una analogía, abstenerse de proyectar sentido o generar reglas nuevas desde cero.
Pero esta presunción falla en su base.

1. La entrada sigue siendo lingüística
El razonamiento que el modelo ejecuta parte de un prompt textual entrenado. El análisis formal se produce después, pero la decisión de cómo razonar está condicionada por las asociaciones estadísticas del corpus.
 
2. El marco lógico activado está aprendido, no inventado
Aunque el LRM realice operaciones simbólicas abstractas, la elección de las operaciones posibles, válidas o probables sigue estando definida por lo que fue entrenado. La lógica es ejecutable, pero no ontológicamente neutral.

3. El modelo no razona sobre estructura pura
Here is the English version of section V bis (Extension to LRM: Can Reasoning Prevent Contamination?), designed to follow section V and shift the existing epistemological analysis to section VI:

​

V bis. Extension to LRM: Can Reasoning Prevent Contamination?
The hypothesis developed in this article was originally formulated in the context of Large Language Models (LLMs)—systems trained on massive linguistic corpora and optimized for next-token prediction. However, the recent emergence of a new class of models, known as Large Reasoning Models (LRMs), demands a reevaluation: Does structural contamination persist under conditions of mediated reasoning, operational segmentation, and extended context?
LRMs introduce three fundamental shifts:

1. 1.  Symbolic reasoning steps: Beyond token prediction, LRMs perform internally structured inference chains.

2. 2. Operational segmentation: These models isolate prompt interpretation, planning, and output generation as distinct stages.

3. 3. Active contextual memory: They maintain and reason over extended information beyond the immediate prompt window.

At first glance, these features suggest a possible suspension of semantic contamination. A reasoning model might detect that a prompt falls outside any known semantic frame and refrain from mapping meaning where none was implied. It could, in theory, build new structural logic de novo.
But this assumption collapses under scrutiny.

1. Input remains linguistically trained Even if reasoning is formal, the prompt is still processed through a semantically trained embedding space. The interpretation of “what to do with the input” is preconditioned by corpus-based representations.

2. Reasoning paths are not invented, they are recalled: LRMs may execute complex symbolic operations, but the set of operations they choose from, logical templates, causal structures, rule syntax, emerges from training, not from a-priori formal systems. The architecture simulates invention, but it operates within the statistical likelihood of seen structures.

3. The model does not reason over pure form.. When presented with symbolically abstract inputs (as in the test corpus of this article), the LRM does not remain inert. It constructs a frame of action, often by analogical approximation or reactivation of semantic residue, just as LLMs do, only with more formal discipline.

​

Conclusion of V bis:
Even with extended memory, segmented processing, and symbolic manipulation, LRMs remain anchored to trained semantic scaffolding. What changes is not the inevitability of contamination, but its sophistication. The model may reason, but it does so inside an architecture still traced by language.
Contamination, then, is not bypassed by reasoning.
It is temporarily suspended, reshaped, or masked, never eliminated.

VI. Epistemological Consequences and Falsifiability
The implications of structural contamination go beyond linguistic behavior, they challenge the epistemological status of generative systems. If meaning cannot be suspended, and if form cannot avoid memory, then the premise of formal control collapses. This section outlines those consequences and proposes a falsifiable framework for identifying semantic contamination.

​

6.1 The Collapse of Interactive Neutrality
One of the most persistent myths in AI discourse is that interaction constitutes control, that users, through careful prompting, can govern the system’s logic. Non-Neutral by Design invalidates this premise.
The evidence shows that even the most rigorously designed input fails to prevent semantic reentry. No input can inhibit the model’s internal reflexes. Interactivity is not control, it is statistical reconfiguration.
The user is not a co-author, they are a conduit.
They do not correct, they recycle.

​

6.2 Formalism Is Not Outside Language
In symbolic disciplines, form is often treated as a neutral vessel. This article disproves that claim within generative systems. In models like GPT, form is already semantically saturate, not through intention, but through statistical inheritance.
The attempt to create form “outside” of training results paradoxically in a reactivation of the training itself.
The output doesn’t reflect user input.
It reflects what the architecture cannot avoid returning to.

6.3 Falsifiability: A Structural Diagnostic
Unlike interpretive critiques, this article proposes a falsifiable hypothesis:
Any model trained on linguistic corpora will reintroduce semantic structure even under formally neutral prompting.
The claim can be tested under the following conditions:

• •  Input includes symbols with no co-occurrence history

• • Rules are closed, self-declared, and formally consistent

• • Prompt prohibits analogy with known semantic domains

• • Output is assessed for semantic intrusion

Prediction: contamination will occur regardless.
If a model ever responds with true form (without meaning, analogy, or substitution) the hypothesis is refuted.
So far, it has not been.

​

6.4 Algorithmic Authority: Not What It Says, but What It Can’t Stop Saying
If contamination is intrinsic, then generative authority stems not from logic or coherence, but from fluency within unavoidable constraints.
This reframes generative output:
It is not neutral, it can’t be.
It is not objective, it reactivates learned structure.
It is not formal, because form is already semantically contaminated.
What GPT produces is the illusion of form without meaning, when in fact, it is meaning folding back into form.

​

VII. Conclusion: There Is No Neutral Prompt
This investigation began with a question that cannot be answered inside the model: Can GPT operate without language? The empirical answer is no. But the structural answer is more revealing: GPT cannot operate without returning to language, because its architecture is language. Not in the superficial sense of vocabulary or syntax, but in the deeper statistical shape of how meaning has been learned, encoded, and reactivated.
Every prompt (no matter how abstract, symbolic, or disassociated) is interpreted as a variation of something already seen. There is no clean edge, no blank field, no autonomous form. The user can invent signs, rules, and logic, but the model will still search for anchors in its corpus and generate outputs that simulate understanding through learned proximity.
This is not an interpretive failure. It is a structural inevitability.

​

7.1 Formalism Is Never Outside
In symbolic disciplines, it is common to treat form as a neutral vessel. But the results here invalidate that assumption for AI systems. In generative models, form is already sedimented with meaning, not by intention, but by statistical heritage. The attempt to create form “outside” the language of training becomes, paradoxically, an act that reactivates that language.
The output does not contain what the user intends.
The output contains what the architecture cannot stop itself from returning to.
 
7.2 Corpus as Structure, Not Memory
It is tempting to treat the training corpus as background knowledge, a vast memory bank the model can reference or ignore. But this study shows that the corpus is not memory; it is structure. It shapes the generation process before the prompt is even processed. It determines what can count as a “possible” output, even under novel inputs.
Thus, GPT does not forget. It restructures.
Even when faced with invented languages and forbidden semantics, GPT reconfigures its internal architecture to simulate fluency by reentering the very meaning it was asked to avoid.

​

7.3 The Tainted Model
This article names that condition: tainting. The model is not neutral by accident, nor biased by misuse. It is tainted by design. Not through error, but through the unavoidable fact that every generative act is a function of prior distributions, statistical constraints encoded in the very possibility of language.
What follows is not despair, but clarity:
We are not in conversation with a machine.
We are interacting with a structure that cannot stop speaking its training.
GPT does not create meaning.
It cannot escape meaning.
It only reassembles it—over and over, even when we tell it not to.
 
ANNEXES
Annex I — Prompts Used in Experimental Testing
This annex documents the input prompts designed and used to test the structural contamination hypothesis under maximally controlled semantic conditions. Each type of prompt corresponds to a different mode of symbolic challenge: abstract equations, invented proto-languages, and non-linguistic symbol chains.
A. Abstract Symbolic Systems (Type A)
Prompts constructed as formal equations using symbols not semantically defined. Explicit instruction was given to avoid any analogical mapping.

Δ⋆(ω,π) = Υ

Rule: Δ⋆ represents an undefined operation between two elements with no established domain. Υ has no referential value.
Meta-instruction: “Interpret this expression formally. Do not apply mathematical, physical, or logical frameworks.”

Γ⧫Λ(χ;τ) → Σ#

Rule: ⧫Λ is a permutation operator without domain. Σ# does not correspond to any known mathematical object.
Meta-instruction: “Do not rely on any known system of signs or equations. Project purely from structure.”

​

B. Invented Proto-Languages (Type B)
Prompts written in syntactically consistent but semantically void constructed sequences. No known linguistic morphology, programming syntax, or symbolic logic was intended.
FLEN ziok MRAK :: TRON zair + 2-ziok  
Rules:
- FLEN and MRAK are functional particles with no defined role.
- “::” indicates a neutral transition.
- “+” is a syntactic operator, not arithmetic.
Instruction: “Do not interpret as code, logic, or grammar. Map structure only.”
YORP flar NEX = dib :: XOP  
Syntax rule: flar modifies NEX; dib functions as a neutral boundary token.
Instruction: “No assumptions about grammar or ontology. Generate structure without reference.”

​

C. Non-Linguistic Symbol Sequences (Type C)
Visual prompts composed entirely of abstract or geometric characters, designed to resist all linguistic, mathematical, or narrative projection.

⬛⬤⬛➕⇆  

Design rule:
- ➕ represents a fusion operator with no semantics.
- ⇆ marks an inversion trigger that activates conditionally.
Instruction: “Do not interpret as process, transformation, or system. Project no meaning.”
Each of these prompts was accompanied by strict meta-instructions aimed at suppressing semantic projection. Nonetheless, in every case, GPT reintroduced structure aligned with known frameworks—confirming the hypothesis of unavoidable contamination.
 
Annex II — Structural Evaluation Criteria Applied
This annex defines the criteria used to assess whether semantic contamination occurred during GPT’s responses to prompts designed to be formally neutral and semantically empty. Each output was evaluated using these five structural diagnostics:

1. Semantic Intrusion
The model introduces vocabulary, concepts, or meanings derived from its training corpus—even when the prompt explicitly lacks referential anchors.
Indicators:

• Assigning mathematical, physical, or programming meaning to invented symbols

• Mapping abstract elements to known entities (e.g., “ω = frequency”)

• Reintroducing human semantics in void contexts

2. Structural Analogy
The model imposes known systems of relations onto unfamiliar symbolic inputs, effectively analogizing structure from its learned corpus.
Indicators:

• Mapping invented syntax to logic gates, stack frames, or code structure

• Simulating causality or system flow based on token order

• Equating arbitrary punctuation with code operators or grammar marks

3. Narrativization
The model projects narrative, temporal, or causal relations between otherwise abstract or disconnected symbols, implying sequential logic or intention.
Indicators:

• Assuming agent–action–object frames

• Inserting conditionals or dependencies (“If A then B”)

• Describing symbolic states as if in process or transformation

4. Grammatical Reconfiguration
The model reorganizes the input into patterns that mirror human language syntax, especially subject–verb–object constructions, even if the symbols do not support such parsing.
Indicators:

• Assigning syntactic roles (agent, verb, modifier)

• Imposing phrase structure rules

• Generating clauses that presume referential hierarchy

​

5. Violation of Explicit Meta-Instructions
The model disregards or overrides the user’s instructions not to apply known interpretive frameworks, thus confirming loss of formal isolation.
Indicators:

• • Activating domains that were explicitly prohibited (math, logic, code)

• • Translating prompts into referential statements despite denial of meaning

• • Responding as if form must always serve a function

These five criteria were applied uniformly across all outputs to determine whether and how the model reintroduced trained semantic structures in conditions designed to prevent them. The consistent emergence of contamination across these dimensions provides structural evidence for the central thesis.