Paper deep dive
Natural Language Declarative Prompting (NLD-P): A Modular Governance Method for Prompt Design Under Model Drift
Hyunwoo Kim, Hanau Yi, Jaehee Bae, Yumin Kim
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 93%
Last extracted: 7/20/2026, 9:52:36 AM
Summary
This paper introduces Natural Language Declarative Prompting (NLD-P), a modular governance method for prompt design that addresses 'GPT-scale model drift.' NLD-P separates prompt components into four invariants: Provenance, Constraint Logic, Task Content, and Post-Generation Evaluation. This structural separation allows for stable, interpretable control and localized revision without relying on external orchestration code, addressing the instability caused by evolving LLM alignment and instruction-following policies.
Entities (8)
Relation Signals (7)
NLD-P → consistsof → Task Content
confidence 98% · NLD-P rests on four structural invariants... Task Content: The primary objective...
NLD-P → consistsof → Provenance
confidence 98% · NLD-P rests on four structural invariants... Provenance: Operational context...
NLD-P → consistsof → Constraint Logic
confidence 98% · NLD-P rests on four structural invariants... Constraint Logic: Behavioral rules...
NLD-P → consistsof → Post-Generation Evaluation
confidence 98% · NLD-P rests on four structural invariants... Post-Generation Evaluation: Conditions for validation...
NLD-P → addresses → GPT-scale model drift
confidence 95% · NLD-P is formalized as a modular control abstraction... positioned as an accessible governance framework... for non-developer practitioners operating within evolving LLM ecosystems.
GPT-scale model drift → causedby → LLMs
confidence 90% · As models scale and update across generations, prompt behavior becomes sensitive to shifts... a phenomenon we characterize as GPT-scale model drift.
RLHF → contributesto → GPT-scale model drift
confidence 85% · Post-training alignment procedures such as reinforcement learning from human feedback (RLHF) demonstrate that substantial behavioral changes can occur... GPT-scale drift stems from revisions to model parameters or safety constraints
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:The rapid evolution of large language models (LLMs) has transformed prompt engineering from a localized craft into a systems-level governance challenge. As models scale and update across generations, prompt behavior becomes sensitive to shifts in instruction-following policies, alignment regimes, and decoding strategies, a phenomenon we characterize as GPT-scale model drift. Under such conditions, surface-level formatting conventions and ad hoc refinement are insufficient to ensure stable, interpretable control. This paper reconceptualizes Natural Language Declarative Prompting (NLD-P) as a declarative governance method rather than a rigid field template. NLD-P is formalized as a modular control abstraction that separates provenance, constraint logic, task content, and post-generation evaluation, encoded directly in natural language without reliance on external orchestration code. We define minimal compliance criteria, analyze model-dependent schema receptivity, and position NLD-P as an accessible governance framework for non-developer practitioners operating within evolving LLM ecosystems. Portions of drafting and editorial refinement employed a schema-bound LLM assistant configured under NLD-P. All conceptual framing, methodological claims, and final revisions were directed, reviewed, and approved by the human author under a documented human-in-the-loop protocol. The paper concludes by outlining implications for declarative control under ongoing model evolution and identifying directions for future empirical validation.
Tags
Links
- Source: https://arxiv.org/abs/2602.22790v1
- Canonical: https://arxiv.org/abs/2602.22790v1
Trouble viewing inline? Open PDF directly →
Full Text
41,970 characters extracted from source content.
Expand or collapse full text
Natural Language Declarative Prompting (NLD-P): A Modular Governance Method for Prompt Design Under Model Drift Hyunwoo Kim Hanau Yi Jaehee Bae Yumin Kim ddai Inc. hw.kim, hnu.yi, jh.bae, ym.kim@ai-dda.com Abstract The rapid evolution of large language models (LLMs) has transformed prompt engineering from a localized craft into a systems-level governance challenge. As models scale and update across generations, prompt behavior becomes sensitive to shifts in instruction-following policies, alignment regimes, and decoding strategies, a phenomenon we characterize as GPT-scale model drift. This paper reconceptualizes Natural Language Declarative Prompting (NLD-P) as a declarative governance method rather than a rigid field template. NLD-P is formalized as a modular control abstraction that separates provenance, constraint logic, task content, and post-generation evaluation, encoded directly in natural language without reliance on orchestration code. Portions of drafting employed a schema-bound LLM assistant configured under NLD-P. All methodological claims were directed and verified under a documented human-in-the-loop protocol. We conclude by outlining implications for declarative control under ongoing model evolution and directions for future validation. 1 Introduction: Prompting as Governance Under Model Drift Large language models (LLMs) have evolved from relatively static generative systems into continu- ously updated architectures shaped by shifting align- ment policies, decoding strategies, and instruction- following heuristics. Post-training alignment proce- dures such as reinforcement learning from human feedback (RLHF) demonstrate that substantial be- havioral changes can occur without modifying base model architecture (Ouyang et al., 2022). Founda- tion model research further situates model behavior within a socio-technical ecosystem in which training data, deployment context, and governance regimes co-produce system outputs (Bommasani et al., 2021). From a sociotechnical perspective, technical artifacts embed forms of governance and constraint within their design (Winner, 1980). Contemporary system documentation confirms that successive model re- leases incorporate iterative safety mitigations, align- ment refinements, and revised refusal policies, re- inforcing that instruction-following characteristics are governed across versions rather than fixed at pretraining (Singh et al., 2025). At the interface level, these governance dynamics are increasingly mediated through natural language. Linguistic analysis indicates that sustained interac- tion with LLM systems gives rise to a hybridized register shaped by system affordances and alignment constraints (Kim & Yi, 2025). Termed machine- facing English, this register reflects the co-adaptation of human expression to model behavior and model alignment to recurring prompting patterns (Kim & Yi, 2025). As model generations introduce behav- ioral adjustments through alignment and deployment updates, prompts once treated as stable artifacts be- come sensitive to architectural change. Under these conditions, prompt engineering must be reconcep- tualized not as phrasing craft but as governance enacted within an evolving linguistic interface (Kim & Yi, 2025). 1.1 The Escalation of Prompt Control Early prompt practices relied on role assignment, it- erative rephrasing, or example-based steering. These approaches implicitly assumed behavioral continuity across model updates. Empirical findings demon- 1 arXiv:2602.22790v1 [cs.CL] 26 Feb 2026 strate, however, that scaling and alignment tuning significantly influence reasoning style, verbosity, and adherence to declared constraints (Ouyang et al., 2022; Wei et al., 2022). Even minor variations in prompt framing or example ordering can alter per- formance characteristics (Min et al., 2022). Frontier model documentation further shows that refusal be- havior and constraint handling are adjusted between releases as part of iterative deployment processes (Singh et al., 2025). Simultaneously, linguistic evidence suggests that users adapt prompting strategies in response to per- ceived system sensitivities, gradually internalizing alignment-aware phrasing patterns (Kim & Yi, 2025). This co-adaptive dynamic intensifies the control bur- den on language itself. As models evolve through post-training and policy updates, identical prompts may vary in constraint enforcement, structural for- matting, or interpretive emphasis. Such variability reflects architectural and governance change rather than user inconsistency. Consequently, prompt design shifts from stylistic op- timization to structural control. Effective prompting increasingly requires explicit constraint visibility and modular logic that can be revised without rewriting entire instruction blocks. Without such separation, behavioral drift accumulates and undermines repro- ducibility across model versions. From a governance perspective, prompt authorship resembles reflective practice under conditions of uncertainty, where de- sign and revision form a continuous feedback loop (Sch ̈on, 1983). Prompt control becomes structured oversight within a non-stationary technical system rather than rhetorical refinement alone. Because gov- ernance is expressed linguistically, structural clarity must counteract the tendency of machine-facing reg- isters to absorb constraints into diffuse narrative phrasing (Kim & Yi, 2025). 1.2 GPT-Scale Model Drift We define GPT-scale model drift as cross-version variability in instruction-following behavior result- ing from architectural updates, alignment modifica- tions, or inference-mode differences. Unlike classi- cal distribution shift, which arises from changes in input data distributions (Quionero-Candela et al., 2009), GPT-scale drift stems from revisions to model parameters or safety constraints between releases. Concept drift literature addresses non-stationary en- vironments in which predictive relationships evolve over time (Gama et al., 2014); GPT-scale drift differs insofar as internal policy updates are intentionally introduced through model revision cycles. Public documentation confirms that alignment refinements and safety policies are revised across releases, in- troducing version-dependent variation in constraint handling (Singh et al., 2025). Drift may manifest as inconsistent enforcement of declared constraints, reinterpretation of task bound- aries, changes in verbosity, or altered adherence to post-generation conditions. Safety research has em- phasized that objective misalignment and distribu- tional fragility can arise even in high-performing systems (Amodei et al., 2016). From a systems per- spective, any shift in constraint behavior introduces instability into externally authored control logic. When governance is encoded in natural language, drift interacts with register-level adaptation. As models shift through updated alignment regimes, users may adjust phrasing to compensate, reinforc- ing the co-evolution of linguistic form and system be- havior (Kim & Yi, 2025). Drift therefore operates at both architectural and interface levels, where gover- nance and language co-adapt. The central challenge lies in designing control abstractions that remain in- terpretable and revisable under architectural change rather than assuming behavioral continuity. 1.3The Governance Gap in Natural- Language Interfaces Developer-centric AI pipelines externalize control logic into code, enabling structured orchestration and validation. Such approaches support reproducibility but require technical infrastructure and program- ming fluency. In contrast, many LLM users oper- ate exclusively through natural-language interfaces, where governance must be encoded linguistically within the prompt itself. Research in end-user programming highlights the difficulty non-developers face when abstract control structures are implicit rather than tool-supported (Ko et al., 2011). Human–AI interaction guidelines similarly emphasize transparency and user control as prerequisites for trustworthy AI systems (Amershi et al., 2019). Concurrently, linguistic analysis indi- cates that as users adapt to evolving model behav- ior, prompting increasingly reflects system-oriented phrasing patterns that prioritize alignment compat- 2 ibility over conversational naturalness (Kim & Yi, 2025). When governance logic is conflated with task in- structions in a single narrative block, users must manage control implicitly within this evolving reg- ister. This increases fragility under model change. Conventional prompting often merges task objec- tives, behavioral constraints, formatting rules, and evaluation expectations into undifferentiated text. Such conflation obscures governance logic and com- plicates targeted revision when drift occurs. Ac- countability frameworks in AI governance advocate lifecycle-based oversight rather than post hoc re- mediation (Raji et al., 2020). In natural-language interfaces, however, such lifecycle governance must itself be expressed in language already shaped by model-mediated conventions (Kim & Yi, 2025). The burden of structural clarity thus shifts decisively to prompt design. 1.4 From Format to Method Natural Language Declarative Prompting (NLD-P) was initially introduced through structured field- based prompts separating identity, rules, and task content. Over time, durability proved to depend not on specific bracket syntax but on a deeper invariant: explicit separation between provenance, constraint logic, task execution, and post-generation evaluation. This structural separation functions as a counter- measure to linguistic drift within machine-facing registers. If governance remains embedded in diffuse narrative phrasing, it becomes vulnerable to reinter- pretation under evolving alignment heuristics and mitigation regimes. By contrast, modular segmen- tation preserves governance clarity even when the surrounding register shifts. This principle aligns with modularity theory in soft- ware architecture, where separation of concerns en- hances system stability and maintainability (Parnas, 1972; Shaw & Garlan, 1996). In complex systems theory, hierarchical decomposition allows local modi- fication without systemic collapse (Simon, 1962). By externalizing governance layers from execution logic, modular systems enable targeted revision rather than cascading restructuring. This paper formalizes NLD-P as a declarative gover- nance method rather than a formatting convention. Its core contribution lies in the externalization of con- trol logic into modular, legible natural-language com- ponents that remain interpretable under model drift. The objective is not to eliminate drift but to render it visible and tractable. When governance layers are structurally separated from task content, behav- ioral deviations can be addressed through localized revision of constraint blocks rather than wholesale prompt reconstruction. The following sections formalize the invariants of declarative modular prompting, define minimal com- pliance criteria, and analyze schema receptivity across evolving LLM architectures. NLD-P is posi- tioned not as a performance optimization technique but as a governance abstraction designed for struc- tural stability under GPT-scale model drift and lin- guistic co-adaptation in human–AI discourse (Kim & Yi, 2025). 2 Declarative Governance With- out Code NLD-P is formalized as a declarative governance method for prompt construction under conditions of model drift. Its central innovation lies in modu- lar separation of instruction layers within natural language itself. Governance logic is made explicit, in- spectable, and revision-ready without reliance on ex- ternal scripting, orchestration frameworks, or weight modification. Whereas developer-centric pipelines externalize control into code, NLD-P embeds gover- nance directly within linguistic structure. This formulation draws conceptually from separation- of-concerns principles in software architecture, where isolating functional components enhances system stability and maintainability (Parnas, 1972; Shaw & Garlan, 1996). It also parallels the distinction between declarative and procedural paradigms in programming theory, in which constraints are artic- ulated as explicit conditions rather than embedded within execution flow (Dijkstra, 1976). In NLD-P, modularity is achieved not computationally but lin- guistically: structure is enacted through segmen- tation of control and task layers within natural- language prompts. 2.1 Core Invariants of NLD-P NLD-P rests on four structural invariants. A prompt aligns with the method when these elements are explicitly separated and independently interpretable: 3 •Provenance: Operational context, role assump- tions, or scope conditions relevant to execution. • Constraint Logic: Behavioral rules, formatting requirements, or boundary conditions declared independently of the task objective. •Task Content: The primary objective isolated from governance directives. • Post-Generation Evaluation: Conditions for validation, revision, or compliance articulated prior to execution. These invariants externalize governance. Rather than embedding expectations implicitly within a single narrative instruction, NLD-P distinguishes between what the model should do and how its be- havior should be constrained or evaluated. This structural separation mirrors modular design theory, where isolating components reduces cascading insta- bility when one component changes (Parnas, 1972). By disentangling execution logic from governance logic, prompts become structurally inspectable, diag- nosable, and locally revisable under changing model conditions. 2.2 Schema Versus Surface Syntax Early implementations of NLD-P employed visible field markers to reinforce structural separation. How- ever, the durability of the method does not depend on specific labels, brackets, or markup conventions. The essential requirement is structural modularity rather than syntactic uniformity. A prompt conforms to NLD-P when constraint logic remains distinct from task content, governance in- structions are independently revisable, and evalu- ation conditions are declared rather than implied. Surface syntax may vary across platforms, interfaces, or stylistic conventions; structural separation, by contrast, preserves governance clarity under behav- ioral change. As architectural theory emphasizes, robustness derives from conceptual layering rather than syntactic representation (Shaw & Garlan, 1996). NLD-P is therefore defined at the level of structural invariants rather than template rigidity. 2.3 Minimal Compliance Criteria For analytical clarity, a prompt is considered NLD-P compliant if it: •Explicitly separates control logic from task instruc- tions. •Encodes behavioral constraints in natural lan- guage prior to execution. • Articulates post-generation validation or revision conditions. • Supports targeted revision without collapsing gov- ernance and task layers. These criteria are descriptive rather than prescrip- tive. They define alignment with the method with- out requiring adherence to a single canonical schema. Compliance is architectural rather than syntactic: what matters is the persistence of modular bound- aries rather than their formatting. 2.4 Reference Schema Illustration To clarify how these invariants may be instantiated, Table 1 presents the abstract structural layers of a canonical NLD-P configuration. The representa- tion is illustrative rather than normative. Block names function as authoring conventions and are not machine-parsed metadata. Compliance depends on preservation of structural separation rather than specific surface labels. A compact schematic instantiation is shown in Fig- ure 1. This illustration demonstrates structural sep- aration rather than prescriptive syntax. Layered separation encodes governance logic directly within natural language. The model does not in- terpret these labels as executable metadata; their function is authorial transparency and lifecycle trace- ability. The abstraction resides in the persistence of modular boundaries rather than in specific markers. 2.5Declarative Modularity as Control Abstraction Traditional developer-centric pipelines externalize governance through code, validation scripts, or or- chestration frameworks. NLD-P instead achieves modularity within the prompt itself. While external tooling may enhance scalability, the core governance abstraction remains embedded in natural language. Declarative modularity therefore functions as a con- trol abstraction. By isolating governance compo- nents, prompts become structurally transparent and revision-ready. When behavior shifts, authors can adjust constraint blocks without rewriting task logic. This localized adjustability reflects classical mod- ular system design, in which abstraction enables 4 BlockFunctional RoleExample Elements IDENTProvenance and execution context schema, assistant, mode, type, version RULEConstraint logic layertone, clarity flags, formatting constraints CONTENTTask execution layerPrimary objective or instruction INPUTOptional contextual groundingReference text or external material POSTCHECK Post-generation evaluationValidation rules, revision triggers Table 1: Canonical NLD-P Structural Layers [IDENT] schema = NLD-P assistant = Evalyn mode = essay_proofreader type = declarative_guidance version = 2.0 [RULE:OUTPUT_STYLE] tone = formal style = academic enforce_clarity = true [CONTENT] Proofread the paragraph and identify ambiguity. [POSTCHECK] If vague modal verbs are detected, flag for revision. If citation format deviates from APA, suggest correction. Figure 1: Illustrative Canonical NLD-P Instantiation local modification without systemic collapse (Par- nas, 1972; Shaw & Garlan, 1996). The significance of NLD-P lies not in template rigid- ity but in abstraction of governance into legible components. Under GPT-scale model drift, explicit structural separation provides a stable conceptual scaffold even as model behavior evolves. 3Prompt Lifecycle as Structured Governance Under model drift, prompt design must be treated as an iterative governance cycle rather than a one-time instruction. NLD-P structures this cycle into four stages: authoring, execution, evaluation, and revi- sion. Each stage remains modular, enabling targeted adjustment without collapsing the overall architec- ture. This lifecycle orientation aligns with models of reflective practice in which design, action, and re- assessment form a continuous loop under conditions of uncertainty (Sch ̈on, 1983). It also parallels double- loop learning frameworks, where underlying govern- ing variables—not merely surface behaviors—are ex- amined and revised (Argyris & Sch ̈on, 1978). Rather than viewing prompting as a static artifact, NLD-P conceptualizes it as managed governance embedded within evolving system behavior. 3.1Authoring with Constraint Visibility Authoring begins with explicit declaration of control logic prior to execution. Behavioral boundaries, for- matting requirements, and evaluation conditions are articulated independently of the task objective. This separation clarifies intent before interpretation and preserves editability: constraints may be strength- ened, relaxed, or reformulated without modifying core task logic. Constraint visibility therefore func- tions as pre-execution governance. 5 From a systems perspective, making control condi- tions explicit increases interpretability and reduces ambiguity in downstream evaluation. Human–AI interaction research identifies transparency and user control as prerequisites for reliable oversight in AI- mediated systems (Amershi et al., 2019). By iso- lating constraints in advance, authoring becomes structured boundary-setting rather than post hoc correction. The prompt shifts from a monolithic instruction to a layered governance specification. 3.2 Execution Under Drift Conditions Execution unfolds within model environments that vary across versions, deployment contexts, or infer- ence modes. Even when instructions remain un- changed, interpretation of constraints may shift due to alignment updates or scaling effects (Ouyang et al., 2022; Wei et al., 2022). In monolithic prompts, such shifts are difficult to diagnose because task instructions and governance logic are intertwined. Modular separation enables behavioral variation to be traced to specific control layers. If adherence weakens, revision can target the constraint bound- ary rather than the task narrative. Execution is thus reframed not as an opaque generation event but as a structural stress test of governance clarity. This framing aligns with lifecycle accountability ap- proaches in AI governance, where system behavior is evaluated relative to explicitly declared policies rather than implicit expectations (Raji et al., 2020). Drift becomes analyzable rather than anecdotal. 3.3 Declarative Evaluation Evaluation is defined prior to execution. The prompt may declare explicit conditions governing acceptable outputs, including formatting requirements, prohib- ited categories, structural constraints, or verification rules. By articulating evaluation criteria ex ante, NLD-P embeds oversight within the design stage itself. When violations occur, corrective logic can be revised directly within the governance layer. Evaluation be- comes intrinsic to the lifecycle rather than appended externally. This approach parallels documentation- based accountability frameworks emphasizing ad- vance specification of evaluation standards and trans- parent reporting practices (Mitchell et al., 2019). Declarative evaluation transforms prompt refinement from reactive adjustment into structured compliance review. Rather than revising outputs ad hoc, authors revise governance specifications. 3.4 Revision and Stability Revision targets discrete governance components. Because provenance, constraints, task content, and evaluation logic are structurally separated, iteration does not require full prompt reconstruction. Adjust- ments can be localized to constraint blocks, eval- uation criteria, or contextual assumptions without destabilizing the task layer. This targeted revision supports stability under drift. The method does not eliminate variability but re- duces cascading instability produced by conflated instruction structures. Stability emerges from preser- vation of governance boundaries rather than assump- tion of invariant model behavior. Lifecycle gover- nance therefore transforms prompting from reactive trial-and-error into structured iteration. Continuity resides in architectural separation, even as model behavior evolves. 4Schema Receptivity and Model Dependency Declarative modular prompting operates within model architectures that differ in how they inter- pret instruction layers. Even when governance logic is made explicit, its effectiveness depends on how a model parses constraint boundaries relative to task content. Research on large language model scaling and behavioral predictability shows that responses shift systematically with training scale and align- ment configuration (Ganguli et al., 2022). Studies of alignment regimes further demonstrate that consti- tutional and reinforcement-based approaches yield materially different patterns of constraint adherence and refusal behavior (Anthropic, 2023). Prompt sensitivity research indicates that structural scaffold- ing and reasoning decomposition can substantially alter outputs even when task content remains con- stant (Yao et al., 2022). System-level documentation across frontier deployments further illustrates that mitigation layering and refusal policies evolve over time, introducing version-dependent variation in con- straint enforcement. This interaction between structural prompt design and model-specific interpretation introduces what we term schema receptivity : the degree to which a 6 given model recognizes, respects, and preserves de- clared governance separations. Schema receptivity is not binary but gradient, shaped by training distribu- tions, alignment objectives, safety filters, moderation pipelines, and decoding heuristics. Governance logic expressed in natural language is therefore mediated through evolving alignment regimes rather than fixed syntactic parsing. 4.1Model-Dependent Constraint Adher- ence Different LLM generations exhibit varying sensi- tivities to declared constraints. Some models ad- here closely to explicit formatting or behavioral boundaries, whereas others reinterpret them through broader alignment heuristics or safety policies. Em- pirical red-teaming research demonstrates that mod- est variations in prompt framing can expose sub- stantial differences in safety enforcement and bound- ary interpretation (Perez et al., 2022). Experimen- tal work on in-context learning further shows that demonstration structure and example ordering sig- nificantly influence adherence patterns (Min et al., 2022). Alignment methodology itself affects constraint be- havior. Constitutional alignment approaches modify how models internalize normative guidelines rela- tive to alternative post-training methods (Anthropic, 2023). Such differences indicate that constraint ad- herence is mediated not only by base architecture but also by post-training governance layers. This variability does not invalidate declarative gov- ernance. Rather, it underscores that instruction- following capacity is neither static nor homogeneous across model iterations. Schema receptivity acknowl- edges that governance abstractions operate within evolving policy regimes. Declarative modularity does not assume invariant compliance; instead, it pro- vides structural visibility when compliance fluctu- ates. When adherence weakens or shifts, modular separation allows authors to reinforce, clarify, or recalibrate governance blocks without altering task content. 4.2Stabilization Through Explicit Con- trol Surfaces Declarative constraint blocks function as identifiable control surfaces. They provide discrete loci for ad- justment when behavior shifts across model versions. Instead of embedding expectations diffusely within narrative phrasing, governance directives remain isolated, inspectable, and revisable. This reflects modular system design principles in which isolating components reduces propagation of unintended side effects during modification (Parnas, 1972). In contemporary deployment practice, alignment refinements and mitigation updates are layered at both model and system levels. Explicit governance blocks within prompts operate analogously at the interface layer, enabling targeted revision when align- ment regimes shift without requiring wholesale re- construction of task instructions. Empirical analyses of scaling and predictability suggest that even when aggregate capabilities stabilize, constraint interpre- tation may continue to vary at the interaction layer (Ganguli et al., 2022). Stabilization in this context does not imply invari- ant outputs across model generations. Rather, it reflects increased responsiveness to structured revi- sion. When constraints are explicitly articulated, deviations can be localized and addressed systemati- cally. Drift becomes diagnosable within prompt ar- chitecture rather than obscured within entangled in- structions. Schema receptivity therefore determines not whether governance universally succeeds, but whether structural separation remains interpretable and actionable under changing model conditions. 4.3Portability and Deployment Context NLD-P is defined at the level of architectural ab- straction rather than platform-specific tooling. Its principles apply across hosted interfaces, API envi- ronments, and research workflows. Governance logic remains embedded in natural language rather than external orchestration frameworks. Interface-level governance is increasingly recognized as a primary site of control in foundation model ecosystems (Bom- masani et al., 2021). Portability is conceptual rather than syntactic. Structural separation preserves interpretability even as deployment contexts, inference modes, or model versions change. Constraint adherence may vary according to schema receptivity, yet the governance abstraction itself remains intact. By maintaining explicit boundaries between task execution and be- havioral control, NLD-P supports cross-context inter- pretability under architectural evolution and model dependency. 7 5Reference Implementation and Transparency The formalization of NLD-P emerged through sus- tained application in drafting, analytical refinement, and governance-oriented prompt design workflows. Because large language models can influence struc- ture, phrasing, and argumentative flow, explicit de- lineation of system assistance and authorship bound- aries is required. Transparency in documentation and attribution has become central in responsible AI research, particularly in relation to model re- porting, dataset documentation, and disclosure of system capabilities (Mitchell et al., 2019; Gebru et al., 2021). Broader critiques of large language mod- els further emphasize epistemic responsibility and the risks of obscuring system mediation in knowl- edge production (Bender et al., 2021). Emerging scholarly discourse on generative AI and authorship similarly stresses that accountability for intellectual claims must remain with human contributors. This section clarifies the operational role of the LLM assistant and the limits of its contribution within the present manuscript. Declarative modular prompt- ing functions not only as a design method but also as a transparency instrument, rendering the bound- ary between governance logic and execution visible within authorship practice. 5.1 Evalyn as a Schema-Bound Drafting Agent An LLM assistant configured under the NLD-P framework was used as a structured drafting instru- ment. It operated under explicitly declared gover- nance constraints, including role definitions, citation discipline, structural limits, and stylistic parame- ters embedded within the prompt schema. These constraints were articulated prior to generation and functioned as boundary conditions governing output scope. The assistant did not introduce independent research claims nor autonomously define conceptual frame- works or theoretical positions. All substantive claims were determined by the human author. Generated drafts were treated as provisional artifacts subject to review, revision, or rejection. The system functioned as a reference implementation of declarative modular prompting rather than as an experimental subject or intellectual collaborator. Its role was to execute schema-bound instructions within a human-directed governance structure. 5.2 Human-in-the-Loop Finalization All conceptual framing, definitional boundaries, and scope limitations were determined by the human author. Draft outputs were reviewed iteratively, edited for conceptual precision, and revised to align with the intended theoretical structure. No section was included without explicit approval and refine- ment. Oversight was continuous rather than post hoc, reflecting lifecycle governance principles consis- tent with accountability frameworks in AI system development (Raji et al., 2020). Constraint blocks were themselves revised across drafting cycles to maintain clarity and alignment with architectural claims. Human supervision gov- erned both content selection and structural modifica- tion. The LLM functioned as a drafting instrument within a controlled declarative environment rather than as a source of autonomous intellectual contri- bution. 5.3 Authorship and Responsibility The intellectual content of this manuscript remains the responsibility of the human author.LLM- assisted drafting does not alter authorship attribu- tion or epistemic accountability. Transparent dis- closure of system involvement aligns with documen- tation practices such as model cards and dataset reporting standards emphasizing clarity of responsi- bility and system capabilities (Mitchell et al., 2019; Gebru et al., 2021). By explicitly defining the assistant’s operational scope and maintaining continuous human oversight, the implementation preserves methodological in- tegrity while avoiding ambiguity regarding intellec- tual ownership. Declarative governance thus func- tions not only as a prompt design abstraction but also as a transparency mechanism, ensuring that authorship, accountability, and system assistance remain structurally differentiated. 6 Limitations and Scope This paper formalizes NLD-P as a declarative gov- ernance method under GPT-scale model drift. It does not claim empirical superiority, performance 8 gains, or universal stability across model architec- tures. The contribution is architectural rather than benchmark-driven, situating NLD-P within broader discussions of alignment-sensitive interface design and foundation-model governance (Bommasani et al., 2021). Contemporary research demonstrates that alignment configurations and post-training regimes materially influence model behavior (Ganguli et al., 2022; Anthropic, 2023). NLD-P addresses this dy- namism at the level of interface abstraction rather than model internals. No quantitative benchmarking is provided. The method is articulated at the level of structural ab- straction rather than experimental comparison. Sys- tematic measurement of constraint adherence across model generations remains future work. Rigorous validation would require controlled cross-version test- ing, clearly defined compliance metrics, and repli- cation across deployment environments sensitive to evolving alignment regimes and inference configura- tions. Declarative modularity does not eliminate variabil- ity. Model-dependent schema receptivity implies that constraint adherence may differ across contexts, inference modes, safety layers, or alignment configu- rations. NLD-P increases visibility and revisability of governance logic but does not guarantee invari- ant execution. Stability derives from architectural separation rather than behavioral uniformity. The reference implementation reflects contemporary deployment environments and current-generation instruction-following models. Operational nuances may vary across future architectures, particularly as alignment methods, tool-augmented reasoning, retrieval integration, or multi-agent orchestration frameworks become more prevalent. As foundation models evolve and governance layers grow more com- plex (Bommasani et al., 2021), interface-level ab- stractions may require adaptation to remain inter- pretable and effective. Minimal compliance criteria are descriptive rather than exhaustive. They define alignment with the method at the level of structural separation without prescribing canonical syntax. Alternative frame- works may satisfy similar invariants under different terminology or representational conventions. NLD-P should therefore be understood as one articulation of declarative modular governance rather than an exclusive formalism. Future work should include empirical evaluation across model generations, measurement of schema receptivity under controlled perturbations, and com- parative analysis with programmatic orchestration approaches, including tool-based pipelines and hy- brid declarative–procedural systems. Longitudinal studies examining how linguistic adaptation interacts with evolving alignment regimes would further clar- ify the relationship between architectural drift and interface-level governance (Kim & Yi, 2025). The present contribution addresses the architectural ques- tion of how governance can be externalized within natural language under model drift, leaving empiri- cal validation to subsequent research. 7 Conclusion As large language models continue to evolve, prompt engineering must be reframed as a governance prob- lem embedded within an adaptive technical and lin- guistic system. Architectural updates, alignment refinements, and deployment shifts alter how con- straints are interpreted over time. When governance is enacted through natural language, these shifts interact with the evolving register of human–AI dis- course (Kim & Yi, 2025). Surface formatting conven- tions and monolithic instruction blocks are therefore insufficient to preserve interpretability under condi- tions of model drift. This paper formalizes Natural Language Declara- tive Prompting (NLD-P) as a declarative modu- lar governance method. Its defining property is explicit structural separation between provenance, constraint logic, task content, and post-generation evaluation. By externalizing control layers into in- dependently revisable components, NLD-P renders drift visible rather than opaque. Governance be- comes inspectable and locally adjustable at the in- terface level, even as alignment regimes and model architectures evolve. NLD-P does not guarantee invariant behavior across systems. Its contribution is architectural: it articu- lates a control abstraction whose coherence does not depend on behavioral uniformity. Structural sepa- ration enables targeted revision without collapsing task logic, allowing governance to adapt alongside evolving models while maintaining conceptual conti- nuity. As users adapt language to model sensitivities, 9 machine-facing registers may absorb governance into diffuse narrative phrasing (Kim & Yi, 2025). Declara- tive modularity counteracts this tendency by preserv- ing the distinction between execution and control. It sustains structural clarity at the intersection of linguistic adaptation and architectural change. The central claim of this work is therefore archi- tectural rather than empirical: when governance is expressed through natural language, modular sepa- ration becomes a design requirement. Under GPT- scale model drift and linguistic co-adaptation, ex- plicit governance segmentation is not stylistic pref- erence but structural necessity for preserving inter- pretability and revisability over time. References [1]Amershi, S., Weld, D., Vorvoreanu, M., Four- ney, A., Nushi, B., Collisson, P., et al. (2019). Guidelines for human–AI interaction. Proceed- ings of the 2019 CHI Conference on Human Factors in Computing Systems, 1–13. [2]Amodei, D., Olah, C., Steinhardt, J., Chris- tiano, P., Schulman, J., & Man ́e, D. (2016). Con- crete problems in AI safety. arXiv:1606.06565. [3]Anthropic. (2023). Constitutional AI: Harmless- ness from AI feedback. arXiv:2212.08073. [4]Argyris, C., & Sch ̈on, D. A. (1978). Organiza- tional learning: A theory of action perspective. Addison-Wesley. [5]Bender, E. M., Gebru, T., McMillan-Major, A., & Shmitchell, S. (2021). On the dangers of stochastic parrots: Can language models be too big? Proceedings of the 2021 ACM Conference on Fairness, Accountability, and Transparency, 610–623. [6] Bommasani, R., Hudson, D. A., Adeli, E., Alt- man, R., Arora, S., von Arx, S., et al. (2021). On the opportunities and risks of foundation models. arXiv:2108.07258. [7]Dijkstra, E. W. (1976). A discipline of program- ming. Prentice Hall. [8]Gama, J., ˇ Zliobait ̇e, I., Bifet, A., Pechenizkiy, M., & Bouchachia, A. (2014). A survey on con- cept drift adaptation. ACM Computing Surveys, 46(4), 1–37. [9]Ganguli, D., Askell, A., Schiefer, N., Liao, T., Knight, H., et al. (2022). Predictabil- ity and surprise in large generative models. arXiv:2202.07785. [10]Gebru, T., Morgenstern, J., Vecchione, B., Vaughan, J. W., Wallach, H., et al. (2021). Datasheets for datasets. Communications of the ACM, 64(12), 86–92. [11]Kim, H., & Yi, H. (2025). Machine-facing En- glish: Defining a hybrid register shaped by hu- man–AI discourse. arXiv:2505.23035. [12]Ko, A. J., Abraham, R., Beckwith, L., Black- well, A. F., Burnett, M., Erwig, M., et al. (2011). The state of the art in end-user software engi- neering. ACM Computing Surveys, 43(3), 1–44. [13]Mitchell, M., Wu, S., Zaldivar, A., Barnes, P., Vasserman, L., Hutchinson, B., et al. (2019). Model cards for model reporting. Proceedings of the 2019 Conference on Fairness, Accountability, and Transparency, 220–229. [14]Ouyang, L., Wu, J., Jiang, X., Almeida, D., Wainwright, C., et al. (2022). Training language models to follow instructions with human feed- back. arXiv:2203.02155. [15] Parnas, D. L. (1972). On the criteria to be used in decomposing systems into modules. Commu- nications of the ACM, 15(12), 1053–1058. [16] Perez, E., Huang, S., Song, H. F., Cai, T., Ring, R., et al. (2022). Red teaming language models with language models. arXiv:2202.03286. [17] Quionero-Candela,J.,Sugiyama,M., Schwaighofer, A., & Lawrence, N. D. (2009). Dataset shift in machine learning. MIT Press. [18]Raji, I. D., Smart, A., White, R. N., Mitchell, M., Gebru, T., et al. (2020). Closing the AI ac- countability gap: Defining an end-to-end frame- work for internal algorithmic auditing. Proceed- ings of the 2020 Conference on Fairness, Ac- countability, and Transparency, 33–44. [19] Sch ̈on, D. A. (1983). The reflective practitioner: How professionals think in action. Basic Books. [20]Shaw, M., & Garlan, D. (1996). Software archi- tecture: Perspectives on an emerging discipline. Prentice Hall. 10 [21]Simon, H. A. (1962). The architecture of com- plexity. Proceedings of the American Philosoph- ical Society, 106(6), 467–482. [22]Singh, A., Fry, A., Perelman, A., Tart, A., Ganesh, A., El-Kishky, A., et al. (2025). Ope- nAI GPT-5 system card. arXiv:2601.03267. [23]Wei, J., Wang, X., Schuurmans, D., Bosma, M., Ichter, B., et al. (2022). Emergent abilities of large language models. arXiv:2206.07682. [24] Winner, L. (1980). Do artifacts have politics? Daedalus, 109(1), 121–136. [25]Yao, S., Zhao, J., Yu, D., Du, N., Shafran, I., Narasimhan, K., & Cao, Y. (2022). ReAct: Synergizing reasoning and acting in language models. arXiv:2210.03629. ... 11