From Language to Computation: Linguistics, Punctuation, Mathematics, and Programming as a Unified:
A
Secretary Suite Project
DOI: to be assigned
John Swygert
August 25, 2026
Abstract
Natural language, punctuation, mathematics, and programming are usually studied as distinct symbolic systems. Large language models (LLMs) provide an operational environment in which these domains meet: linguistic input is tokenized, mapped to numerical representations, transformed mathematically, executed through software and hardware, and returned as generated symbolic output. This paper does not claim novelty for the established facts that punctuation, order, grouping, scope, context, or formatting can alter model behavior, nor for the fact that transformer inference is mathematical. Its contribution is instead a unified relational architecture linking these facts across natural language, mathematical notation, programming syntax, and LLM computation, together with a common perturbational methodology for measuring their effects.
The central distinction is Symbol Identity ≠ Symbol Relation. The same constituents can produce different outcomes when their order, boundaries, scope, hierarchy, context, or historical path changes. The paper develops a dynamic state model of interaction, introduces a relational sensitivity vector, relational compression, minimum symbolic intervention, and a testable form of linguistic hysteresis, and places these constructs within a disciplined progression from mapping to modeling, simulation, prediction, and empirical test. Within TSTOEAO, the paper treats relational mapping as a candidate scientific operation rather than proof: mappability is not validity. The resulting framework provides a foundation for controlled experiments in LLM analysis and suggests future applications to agent reasoning, code diagnosis, symbolic provenance, and shard-based knowledge systems.
1. Introduction
Human language appears fundamentally different from mathematics or computer programming. Words acquire meanings through culture and use; sentences depend on syntax, semantics, pragmatics, discourse, tone, speaker intention, social relation, and prior knowledge. Programming languages are deliberately constrained formal systems. Mathematical notation is designed to reduce ambiguity. Natural language often tolerates and even exploits ambiguity.
Large language models complicate this separation because linguistic material enters a numerical computational process. At a high level:
Human Language → Tokenization → Numerical Representation → Mathematical Transformation → Probability Distribution → Generated Language
The software and hardware that execute these transformations are themselves implemented through formal programming systems. The important point is not that language, mathematics, and programming are identical. It is that they become operationally coupled inside one running architecture.
Language ↔ Punctuation ↔ Mathematics ↔ Programming ↔ Computation
This convergence makes a relational question experimentally accessible: which aspects of symbolic organization actually change the computational trajectory and observable output?
2. Statement of Contribution and Scope
The paper must distinguish established background from the contribution being proposed. Punctuation effects, order sensitivity, scope ambiguity, tokenization, transformer mathematics, prompt-context effects, and delimiter sensitivity are already known subjects across linguistics, NLP, and computer science. The present work therefore does not present those facts as discoveries.
The contribution is the combination of four moves:
A unified cross-domain relational architecture that treats natural-language punctuation, mathematical grouping, programming syntax, and LLM input structure as comparable relational interventions without claiming that the domains are identical.
A common experimental vocabulary for perturbing punctuation, order, boundary, scope, context, and hierarchy while preserving other variables.
A family of operational measures—relational sensitivity, relational compression, minimum symbolic intervention, and relational residuals—intended to turn qualitative observations into measurable model behavior.
A dynamic trajectory account in which conversation is path dependent, permitting explicit tests of linguistic hysteresis and convergent-history effects.
The novelty claimed is therefore one of synthesis, formalization, and experimental program design. Whether any individual metric proves useful is an empirical question.
3. Symbol Identity Is Not Symbol Relation
The recurring principle across the domains considered here is that constituent identity alone is insufficient to determine system behavior.
Symbol Identity ≠ Symbol Relation
Three symbols A, B, and C have the same inventory whether considered as {A,B,C} or {A,C,B}, but their ordered sequences differ:
(A,B,C) ≠ (A,C,B)
Order itself carries information. So do boundaries, grouping, adjacency, hierarchy, scope, and context. A symbolic system therefore requires more than an inventory of components.
Structured Outcome = F(Identity, Order, Boundary, Scope, Hierarchy, Context, History)
4. Punctuation as a Relational Layer
Punctuation is often described as a readability aid or orthographic representation of pauses. Those descriptions are incomplete. Punctuation can alter grouping, attachment, scope, termination, interpolation, attribution, and interpretive expectation.
I = F(W, P, C)
Here W is lexical material, P is punctuation structure, and C is context. Holding W constant while changing P can alter I. This makes punctuation experimentally useful because it can sometimes serve as a low-magnitude intervention with a disproportionately large downstream effect.
A comma can separate; parentheses can embed; a colon can establish a relation between structures; quotation marks can change attribution; a period can terminate one domain and create another. These marks do not merely decorate words. They help define relations among them.
5. Mathematics as an Explicit Relational System
Mathematics makes the same principle visible through operator precedence and grouping.
2 + 3 × 4 = 14
(2 + 3) × 4 = 20
The numbers and operators are unchanged. Their relational organization differs. The constituent set does not uniquely determine the evaluated result.
Constituent Set ≠ Structured Expression
Parentheses, precedence, scope, and order are therefore operational. They determine which transformation is performed.
6. Programming Languages Make Relation Executable
Programming languages extend relational structure into machine behavior. Execution depends on sequence, nesting, scope, branching, delimiters, binding, types, environment, state, and control flow. A small delimiter or indentation error can alter behavior or prevent execution entirely.
B = F(T, R, E)
Here T represents tokens, R relations among them, E the execution environment, and B resulting behavior. Programming makes explicit a point that language often obscures: relations can be executable constraints.
7. The LLM as an Operational Bridge
An LLM can accept prose, punctuation, equations, source code, markup, dialogue, tables, logical operators, and mixed symbolic forms within the same input stream. The model does not require a separate nonmathematical substrate for each domain.
X = (x₁, x₂, …, xₙ)
Tokenization: X → T = (t₁, t₂, …, tₘ)
Representation: tᵢ → eᵢ ∈ ℝᵈ
Layer transformation: H⁽ˡ⁺¹⁾ = Fₗ(H⁽ˡ⁾)
Output: P(tₙ₊₁ | t₁, …, tₙ)
Language enters; mathematical transformations operate; software executes; language returns. The bridge is therefore operational rather than merely metaphorical.
8. Language Is Not Literally Converted into Source Code
A technical distinction is necessary. Natural-language input is not ordinarily compiled into conventional source code before the LLM processes it. The transformation is more accurately represented as:
Language → Tokens → Numbers → Mathematical State Transformations
Programming languages operate at the implementation and infrastructure layers, while token representations and parameters participate in numerical inference. Keeping these layers distinct prevents an unnecessary overstatement while preserving the larger cross-domain connection.
9. Order, Boundary, and Scope as Independent Variables
Order, boundary, and scope should be experimentally separated rather than treated as one vague notion of syntax.
9.1 Order
Xσ = (tσ(1), tσ(2), …, tσ(n))
A sequence sensitivity measure can be written:
ΠO = D(F(X), F(Xσ))
9.2 Boundary
I = F(T, O, B, C)
Boundary structure B determines which elements are interpreted as belonging together.
9.3 Scope
−(a+b) ≠ −a+b
In language, mathematics, and code, an operator applied to one constituent may produce a different outcome from the same operator applied to a larger span. Scope is therefore relational information.
10. Context as a Dynamic Variable
A conversation cannot be modeled adequately as a bag of independent messages. Each new utterance is interpreted from a state created partly by prior interaction.
Sₜ₊₁ = F(Sₜ, Uₜ, Cₜ, Rₜ)
Sₜ is the current interpretive state, Uₜ the new utterance, Cₜ accumulated context, and Rₜ relevant relational configuration. The interaction forms a trajectory:
S₀ → S₁ → S₂ → … → Sₙ
Meaning is therefore not only a property of the current message; it is also a property of the path used to reach the current state.
11. Interpretation as a Probability Landscape
For an LLM, contextual change is observable through changes in probability distributions over continuations. If M denotes candidate interpretations, then context can reweight them:
P(M | Cₜ) → P(M | Cₜ, Uₜ)
A clarification may narrow uncertainty. An ambiguous phrase may preserve multiple candidates. A later sentence may retroactively alter the most plausible interpretation of an earlier one.
This motivates an interpretive potential abstraction Φₜ(M), not as a claim about literal physical potential, but as a mathematical way to represent changing accessibility of candidate meanings.
12. Punctuation as a Controlled Perturbation
Punctuation is especially valuable experimentally because one can hold lexical content constant while introducing a narrow intervention.
X′ = X + ΔP
ΠP = D(F(X), F(X + ΔP))
D may represent divergence in token probabilities, semantic output, activation state, task outcome, generation entropy, or another defined observable. The measurement must be specified before interpretation.
13. Relational Sensitivity Vector
Rather than reducing relational effects to a single robustness score, a model can be characterized by a multidimensional relational sensitivity profile:
Rₛ = [rP, rO, rB, rS, rC, rH]ᵀ
rP: punctuation sensitivity
rO: order sensitivity
rB: boundary sensitivity
rS: scope sensitivity
rC: contextual sensitivity
rH: hierarchical sensitivity
This vector is intended as a measurement framework, not a claim that these six dimensions exhaust linguistic relation. Different tasks and model architectures may require additional dimensions.
14. Relational Compression
A small symbolic intervention can have a large downstream effect. Relational compression measures that leverage:
Cᵣ = Δ Interpretive Structure / Magnitude of Symbolic Intervention
The numerator must be operationalized for a given experiment—for example, output divergence, probability redistribution, task accuracy change, or activation displacement. The purpose is to quantify how much structural consequence is produced per unit of symbolic change.
15. Minimum Symbolic Intervention
A related question is the smallest intervention capable of producing a predefined meaningful effect.
δXmin such that D(F(X), F(X + δXmin)) > θ
The intervention may be one punctuation mark, one moved word, one delimiter, one capitalization change, one scope marker, or one contextual sentence. This creates a controlled way to search for high-leverage relational boundaries.
16. Linguistic Hysteresis
Path dependence suggests a testable hypothesis analogous to hysteresis: two interactions may arrive at apparently similar semantic endpoints by different histories yet respond differently to the same subsequent input.
ΓA: S₀ → … → Sₙ
ΓB: S′₀ → … → S′ₙ
If Sₙ ≈ S′ₙ, test whether F(Sₙ,U) = F(S′ₙ,U)
A systematic inequality would demonstrate residual history dependence. The experiment can be constructed with convergent conversational trajectories followed by an identical probe. This is more specific than the generic claim that context matters.
17. Relational Residual
TSTOEAO places importance on residuals. A linguistic or model-behavior residual can be defined as:
Rᴸ = Yobserved − Ypredicted
A large residual indicates that the proposed relational variables are incomplete, misweighted, or incorrectly specified. Residuals should therefore be treated as diagnostic information rather than hidden as noise.
18. Tokenization as a Confound and Experimental Variable
Tokenization must be given greater methodological weight than a footnote. The same visible string may be segmented differently by different tokenizers, and punctuation can affect token boundaries. An apparent punctuation effect may therefore be partly a tokenization effect rather than a higher-level relational effect.
Experiments should therefore use at least one of three controls:
Hold the tokenizer fixed when comparing relational perturbations within one model family.
Record tokenization for every stimulus so that surface-string changes can be distinguished from token-boundary changes.
Where feasible, repeat matched experiments across different tokenizers and architectures to test whether the effect generalizes.
Tokenizer dependence is not merely a nuisance. It can itself become a measurable layer in the relational pathway.
19. Human and Model Interpretation Must Remain Distinct
An LLM is not a complete model of human cognition. A measurable shift in model probability or activation does not automatically establish the same mechanism in human interpretation. Parallel human studies would be required before extending model-specific findings to human psycholinguistics.
The useful claim is narrower: LLMs provide mathematically instrumentable systems in which linguistic relations can be perturbed and their downstream consequences measured.
20. Cross-Domain Comparative Controls
Mathematics and programming provide useful control domains because relational rules can often be specified exactly. Natural-language cases can then be compared against formal cases in which the correct grouping, scope, or execution is known.
Natural Language → Constrained Language → Mathematical Notation → Programming Syntax
This is not a hierarchy of intelligence or value. It is a continuum of explicitness in relational constraint.
21. Cross-Domain Symbol Function
The same visible symbol can perform different functions in different systems. Parentheses can mark a parenthetical phrase, mathematical grouping, function calls, tuple structure, or precedence. Therefore:
Symbol Form ≠ Symbol Function
Function emerges from relational environment. This creates a useful experimental design: hold the glyph constant while changing the system in which it operates.
22. TSTOEAO Mapping
TSTOEAO is concerned with relational organization, boundaries, pathways, transformations, constraints, equilibrium conditions, and observable outcomes. Language-model behavior provides a concrete domain in which these objects can be operationalized.
ΔBoundary → ΔRelation → ΔState → ΔObservable
Within the canonical relation V = E × Y, the goal is not to force every sentence into a simplistic multiplication. E, Y, and V must be defined for each empirical application, and the mapping must generate testable consequences.
23. Mappability Is Not Validity
Mappability ≠ Validity
Mathematical Representation ≠ Empirical Confirmation
A framework that can retrospectively describe anything is scientifically weak unless it produces constrained predictions, useful compression, or discriminating tests. Relational mapping should therefore be treated as a candidate operation that opens a path to science rather than as proof that the theory has succeeded.
24. Mapping, Modeling, Simulation, Prediction, Test
Observation → Mapping → Formalization → Simulation → Prediction → Empirical Test
Not every mapped phenomenon can presently be simulated, and not every simulation is accurate. Some systems may permit only partial formalization. Those distinctions identify the current scientific step rather than invalidating the relational description.
25. Experimental Program
A practical program can be organized into matched stimulus families.
26. A Stronger Interpretation of Prompting
Prompt engineering can be understood as controlled state preparation. Formatting, examples, delimiters, role assignment, ordering, and constraints create a state before the primary task is presented.
C₀ → C₁ → C₂ → Task Interpretation
This perspective shifts analysis from 'which wording works' toward 'which relational interventions reliably move the system into a desired and measurable state.' It also provides a basis for comparing prompt variants more rigorously.
27. Implications for LLM Agents
The framework has direct implications for LLM agents because an agent does more than generate isolated replies. It maintains context, selects tools, interprets instructions, decomposes tasks, evaluates intermediate results, and updates state over time. A relational lens can therefore be applied to agent planning, instruction hierarchy, tool-call boundaries, memory injection, retrieval context, error recovery, and response construction.
Future work should test whether relational sensitivity profiles predict agent failure modes: for example, whether a model that is unusually sensitive to delimiter changes is also more vulnerable to boundary confusion between user instructions, retrieved text, tool output, and system constraints.
This topic warrants a separate paper because it moves from symbolic analysis into agent architecture, evaluation, and governance.
28. Implications for Code Diagnosis
Programming is a natural extension because software failures frequently arise from relational errors: missing delimiters, incorrect nesting, wrong precedence, scope leakage, order dependence, state mismatch, type mismatch, or malformed interfaces. A relational diagnostic system could decompose a failing program into symbol identity, boundary, order, scope, hierarchy, context, and state, then rank which relational class is most likely responsible for the observed failure.
This does not replace conventional debuggers, static analyzers, type systems, or tests. It provides a cross-language diagnostic lens that may help an LLM reason about why code fails rather than merely regenerate alternatives.
29. Implications for Shard-Based Knowledge Systems
The same relational framework is potentially foundational for shard-based knowledge architectures. A shard can be treated not only as stored content but as a relational unit with provenance, scope, dependencies, frequency of reuse, contextual compatibility, and reconstruction history.
Statistical analysis can then ask which shards are retrieved most often, which combinations recur, which phrases or sentence structures recur identically, which relational templates are overused, and which outputs are reconstructed from unusually similar symbolic pathways.
These measurements could support redundancy analysis, provenance tracking, plagiarism prevention, stylistic diversification, retrieval optimization, and digital fingerprinting. The objective need not be to accuse a source of plagiarism; it can instead be to detect excessive reuse before publication and deliberately reconstruct an output from different relational components.
Because this application involves SecretarySuite and The Shard Library specifically, it should be developed as a separate systems paper rather than overloaded into the present conceptual bridge.
30. From Deconstruction to Reconstruction
The larger methodological value is a lens for deconstruction and reconstruction. A complex symbolic artifact can be decomposed into constituent identity, order, boundary, scope, hierarchy, context, provenance, and history. Those dimensions can then be measured, compared, regrouped, or recomposed.
Artifact → Relational Decomposition → Measurement → Recombination → New Artifact
This suggests a general method for studying why two outputs are similar, why one structure fails, which fragments dominate a reconstruction, and which relationships are carrying most of the functional load.
31. Limitations
LLMs are not complete models of human cognition.
Output behavior does not uniquely identify internal mechanism.
Tokenization can confound apparent punctuation or formatting effects.
Different architectures may implement the same behavioral relation differently.
Cross-domain structural similarity does not imply that natural language, mathematics, and programming are identical systems.
Relational metrics require operational definitions; otherwise equations become decorative notation.
TSTOEAO mappings must be compared with simpler or competing models rather than assumed superior.
Statistical similarity and digital fingerprinting require careful thresholds to avoid false claims of copying or provenance.
32. Methodological Guardrails
Distinguish analogy from mechanism.
Define variables operationally before measuring them.
Separate mapping from validation.
Separate simulation from empirical confirmation.
Preserve lexical content when testing punctuation effects.
Preserve constituent inventory when testing order effects.
Record tokenizer behavior explicitly.
Treat model-specific findings as model-specific unless cross-model replication supports generalization.
Use residual disagreement as information.
Compare TSTOEAO-derived models against simpler alternatives.
33. Unified Relational Expression
Yₜ = F(Tₜ, Oₜ, Bₜ, Sₜ, Cₜ, Rₜ, Hₜ)
Tₜ represents token or symbol identity; Oₜ order; Bₜ boundary structure; Sₜ scope; Cₜ context; Rₜ relational state; Hₜ history; and Yₜ the resulting interpretation or observable output. This is not proposed as a complete language equation. It is a decomposition that makes one claim explicit:
Yₜ ≠ F(Tₜ) alone
34. Conclusion
Large language models reveal an operational convergence of disciplines that are usually siloed. Natural language enters as symbolic input; punctuation and syntax help define relational structure; tokenization converts the sequence into computational units; numerical representations enter a high-dimensional state space; mathematical transformations alter that state; software and hardware execute the transformations; and generated symbolic output returns to language.
Language → Relational Structure → Tokenization → Numerical Representation → Mathematical Transformation → Computational Execution → Language
The central lesson is not that language is 'just mathematics' or that punctuation is equivalent to code. It is that symbolic identity alone does not determine outcome. Order, boundary, scope, hierarchy, context, and history alter the operation performed by the whole.
Symbol Identity ≠ Symbol Relation
The publishable contribution of this paper is therefore not the familiar observation that punctuation or context can matter. It is the attempt to unify these effects across language, mathematics, programming, and LLM computation; to make them measurable through a common perturbational framework; and to distinguish clearly between candidate relational mapping and scientific validation.
Mappability ≠ Validity
For TSTOEAO, this distinction is essential. Relational mapping becomes scientifically valuable only when it supports formalization, simulation where possible, prediction, residual analysis, and empirical discrimination.
Observation → Mapping → Formalization → Simulation → Prediction → Empirical Test
The immediate next step is measurement: controlled experiments on punctuation, order, boundary, scope, context, tokenization, and path dependence across multiple model architectures. Beyond that, the same relational lens can be tested in LLM agents, code diagnosis, provenance analysis, and shard-based knowledge systems. Those applications are substantial enough to warrant separate descendant papers.
The bridge does not need to be invented. It is already running. The task now is to characterize its relational structure, measure its transformations, and test whether those measurements predict anything that simpler descriptions do not.
References
Austin, J. L. (1962). How to Do Things with Words. Oxford University Press.
Brown, T. B., Mann, B., Ryder, N., Subbiah, M., Kaplan, J., Dhariwal, P., et al. (2020). Language models are few-shot learners. Advances in Neural Information Processing Systems, 33, 1877–1901.
Chomsky, N. (1957). Syntactic Structures. Mouton.
Clark, H. H. (1996). Using Language. Cambridge University Press.
Devlin, J., Chang, M.-W., Lee, K., & Toutanova, K. (2019). BERT: Pre-training of deep bidirectional transformers for language understanding. Proceedings of NAACL-HLT 2019, 4171–4186.
Grice, H. P. (1975). Logic and conversation. In P. Cole & J. L. Morgan (Eds.), Syntax and Semantics, Volume 3: Speech Acts (pp. 41–58). Academic Press.
Harris, Z. S. (1954). Distributional structure. Word, 10(2–3), 146–162.
Jurafsky, D., & Martin, J. H. Speech and Language Processing. Stanford University.
Kaplan, J., McCandlish, S., Henighan, T., Brown, T. B., Chess, B., Child, R., et al. (2020). Scaling laws for neural language models. arXiv:2001.08361.
Mikolov, T., Chen, K., Corrado, G., & Dean, J. (2013). Efficient estimation of word representations in vector space. arXiv:1301.3781.
Montague, R. (1974). Formal Philosophy: Selected Papers of Richard Montague. Yale University Press.
Pierce, B. C. (2002). Types and Programming Languages. MIT Press.
Searle, J. R. (1969). Speech Acts: An Essay in the Philosophy of Language. Cambridge University Press.
Shannon, C. E. (1948). A mathematical theory of communication. Bell System Technical Journal, 27, 379–423, 623–656.
Swygert, J. (2026). Punctuation as Linguistic Mathematics. Ivory Tower Publishing.
Swygert, J. (2026). Relational Symbolic Technologies across Language, Mathematics, and Code. Ivory Tower Publishing.
Swygert, J. (2026). TSTOEAO Empirical Core v1.0.0. Ivory Tower Publishing.
Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A. N., Kaiser, Ł., & Polosukhin, I. (2017). Attention is all you need. Advances in Neural Information Processing Systems, 30.
Wittgenstein, L. (1953). Philosophical Investigations. Blackwell.
