From Stored-Program to Intent-Conditioned Computation (IIC)

Foundation-model-based intelligent systems increasingly use natural-language intent as their primary operational interface. In such systems, behavior is conditioned by high-level goals interpreted by learned models at runtime, rather than being fully specified in deterministic source code. This shift creates a systems-level gap that is not fully addressed by either classical software engineering or machine learning theory: how should we design, constrain, evaluate, and govern systems whose behavior emerges from model inference over underspecified intent? This paper proposes Intent-Conditioned Computation (ICC) as a systems-level framework for analyzing and engineering such systems. ICC identifies three structural layers: an intent interface, a model policy layer, and a deterministic infrastructure layer. These layers are connected through an explicit gatekeeper mechanism and an Intent–Action–Observation–Adaptation (IAOA) feedback loop. The framework is grounded in ArchHarness, an open-source enterprise architecture tool that illustrates how stored-program computation and intent-conditioned computation can coexist within a single real-world workflow. The paper derives implications for system design, evaluation, and governance, introduces Realization Fidelity as an ICC-specific evaluation metric, and outlines open research directions for foundation-model-based intelligent systems.