For more than half a century, computing has been understood through the stored-program abstraction: programs encode explicit control flow, and machines execute them deterministically. This model has shaped hardware architecture, programming languages, and software engineering practice. Contemporary AI systems-particularly foundation models and agentic architectures-operate in ways that increasingly strain this abstraction. Behavior is no longer determined solely by fixed code paths; it is conditioned by high-level intent specifications interpreted by large-scale learned models at runtime. This article argues that we are witnessing a shift in computational emphasis, and introduces Intent-Conditioned Computation (IIC) as a complementary abstraction for understanding this shift. In IIC, intent serves as the primary interface, models function as general-purpose policy engines, and execution unfolds as probabilistic, context-sensitive processes. Crucially, IIC does not replace classical computing: modern AI systems stratify both paradigms into complementary layers, with deterministic infrastructure handling reproducibility-critical tasks while intentconditioned models handle flexible, context-sensitive reasoning. We ground this argument in a concrete open-source case study-the ArchHarness enterprise architecture tool-that demonstrates within a single workflow how stored-program and intent-conditioned computation coexist, divide labor, and interact. We discuss implications for system design, evaluation, and governance, and identify five open research directions for the AI community.

Contemporary artificial intelligence systems, particularly large language models (LLMs), are increasingly governed not only by what they know, but by how their behavior is constrained, steered, and revised. Current discourse frequently conflates these governing mechanisms with knowledge itself, leading to conceptual confusion and weakened accountability. This paper argues that such conflation constitutes a fundamental category error. To address this, we propose DIKCA-an extension of the classical DIKW hierarchy that explicitly introduces Control and Autonomy as distinct computational and socio-technical layers. In DIKCA, control specifies permissible and prioritized actions, while autonomy denotes a system's capacity to modify its own control structures over time. We argue that intelligence, in this framework, is best understood as self-modifying control rather than the accumulation of knowledge. By separating epistemic capacity from behavioral governance, DIKCA provides a principled basis for algorithmic accountability, legal responsibility, and democratic oversight. The framework clarifies where failures occur-whether in data, knowledge, control, or autonomy-and enables governance mechanisms that are both technically grounded and socially contestable.

Foundation models are increasingly deployed as execution engines mediating the mapping from symbolic inputs to concrete actions. Traditional metaphors—treating models as programs, interpreters, or knowledge bases—fail to account for their stochastic, latent, and training-dependent execution behavior. We propose Probabilistic Program Execution Semantics (PPES), a formal framework in which foundation models are interpreted as probabilistic interpreters: runtime programs correspond to prompts, execution states include both observable outputs and hidden latent variables, and transitions are governed by a learned Markov kernel K𝜃 defined in the Giry measurable framework. Execution is resolved at infer-time, a distinct semantic phase absent in classical computing. PPES provides a rigorous account of key phenomena in foundation-model-based systems. We develop small-step probabilistic semantics formulated as a sub-probability transition measure (not a set-theoretic relation), define probabilistic reachability over execution traces, and prove four structural theorems: almostsure termination, trace monotonicity, Lipschitz continuity of semantics under parameter perturbation (in total variation, not KL divergence), and the Markov property of latent state. We also establish a notion of 𝜀-observational equivalence on prompts and prove a congruence theorem for sequential composition. Version note. This preprint (v2) revises v1 in four respects: (1) K𝜃 is now formally defined as a Giry-style Markov kernel with explicit measurability conditions; (2) the small-step rule is reformulated to avoid placing a sampling operation in an SOS premise; (3) the smoothness property uses total variation distance with an explicit Lipschitz condition; (4) all four structural properties are now stated as theorems with complete proofs. An extended theoretical development appears in [4]; engineering implications and the ArchHarness case study appear in [5].

Language, traditionally studied from linguistic and cognitive perspectives, can also be formalized as a physical system constrained by energy, entropy, and information. We propose Linguistic Physics, a unified framework interpreting language as an emergent, energy-constrained, and dynamically evolving system. This perspective integrates information theory [1]–[3], statistical mechanics [4], dynamical systems [5], and cognitive neuroscience [6], [7] to model phenomena such as compression, ambiguity, entropy regulation, semantic topology, phase transitions, alignment, evolution, creativity, multimodal emergence, hyper-language structures, and socio-cultural dynamics. We outline theoretical foundations, methodological strategies, and potential contributions, providing a roadmap for computational, experimental, and AI applications.

We introduce a context-dependent probabilistic execution structure strictly extending classical probabilistic transition systems. Unlike Segala systems [14], transition probabilities in our model may depend on execution context, represented as histories. We establish a strict non-embeddability theorem showing that classical probabilistic transition systems cannot capture such context-dependent behaviour. We then construct trace semantics via measurable trace spaces in the sense of Giry [4] and prove a final coalgebra representation theorem following standard coalgebraic methodology [13]. Furthermore, we derive a canonical minimal realisation through trace equivalence, yielding a probabilistic Myhill-Nerode theorem. Finally, we prepare the ground for modal and first-order logical characterisations developed in subsequent sections.

Multi-view architecture descriptions are widely used to manage complexity, yet architecture documentation frequently degrades even when all relevant viewpoints and concerns are covered [12]. Existing coverage-oriented theories fail to explain this phenomenon because they remain silent on where representational authority resides and how relations across views are epistemically maintained. This paper introduces Architectural Responsibility Configuration (ARCM), a responsibility-oriented theory that elevates representational responsibility-the assignment of epistemic authority over architectural facts and relations-to a first-class theoretical primitive, following a theory-building approach in software engineering [28].ARCM yields a closed, falsifiable theory that explains documentation decay and stability as predictable outcomes of responsibility configurations under epistemic gravity and evolutionary pressure.