Knowledge Propagation in AI Engineering and the JDR Self-Descriptive Header: A Proposal for an Entry-Control Mechanism in Natural-Language Job Descriptions

Abstract As generative AI systems increasingly interpret and execute natural-language instructions, documents are no longer limited to explanation or communication; they can become executable media that shape AI behavior. This paper examines this shift through the notion of knowledge propagation in AI engineering. Knowledge propagation refers here to the process by which human intentions, procedures, constraints, state-management rules, and safety policies are transferred through a natural-language document and reflected in an AI system's execution. The paper focuses on the Job Descriptor (JDR), used in yui-AI-OS, as a compact example of a natural-language job description. A JDR differs from a one-shot prompt because it may define a continuing job with dialogue, branching, human confirmation, execution constraints, and improvement records. However, the same document may be interpreted by an AI as a summary target, an editing target, a reference, or an executable job. To reduce this ambiguity, the paper proposes the JDR self-descriptive header as an entry-control mechanism. The header declares document type, intended AI action, execution mode, runtime assumptions, state management, human-gate support, and safety policy. It functions as a meta-knowledge layer that stabilizes the transfer of executable knowledge into AI behavior.