Declarative AI Architecture proposes a knowledge-centric architectural paradigm for generative AI systems. Instead of treating large language models as the primary carriers of domain knowledge and operational behavior, this approach separates generative reasoning capability from explicit knowledge definition. The paper introduces Knowledge Artifacts as persistent, structured, versioned and governable carriers of operational system logic. These artifacts do not merely provide contextual information, as in conventional retrieval-augmented generation. Rather, they define rules, schemas, policies, parameters, resolver logic, validation criteria, source hierarchies, runtime contracts and governance structures that shape how a generative model interprets inputs, constructs outputs and handles uncertainty. Version 2. 4 extends the framework with a more practical and technically precise account of artifact-based system design. It introduces a functional taxonomy of Knowledge Artifacts, grouped into four classes: Domain Logic Artifacts, Structural and Interface Artifacts, Runtime and Activation Artifacts, and Governance, Assurance and Security Artifacts. It also defines an artifact lifecycle for generating, validating, reviewing, binding, testing and maintaining operational knowledge structures. A central addition is the concept of the Artifact Runtime Kernel. Knowledge Artifacts do not execute themselves; they require a stable runtime mechanism that activates, interprets, sequences and enforces the declared artifact environment during inference. The paper distinguishes semantic interpretation by the generative model from deterministic validation by schemas, QA gates, release checks and runtime enforcement mechanisms. This distinction is essential to avoid reducing the approach to prompt engineering or uncontrolled orchestration. The paper positions Declarative AI Architecture in relation to RAG, prompt engineering, agent-based systems, rule-based systems, knowledge graphs and neuro-symbolic AI. It further discusses evaluation criteria, artifact patchability, regression testing, runtime leakage, governance implications and the role of MARAᵥ2 as a reference demonstrator. The core claim is that generative AI systems can become more controllable, inspectable, maintainable and governable when operational knowledge is externalized into structured artifacts and activated through an explicit runtime contract. Declarative AI Architecture does not make generative AI inherently deterministic or error-free; it provides a framework for making its operational behavior more explicit, testable and accountable.
Thomas Gessler (Sun,) studied this question.