From Documents to Semantic Artifacts: A Knowledge-First Framework for Human–AI Work

Documents are effective instruments for communication, interpretation and record, but they compress the journey of work into selected forms. Sources, observations, failed attempts, competing interpretations, decision rationale, expressive context and uncertainty are consequently dispersed across files, conversations, media, tools and memory. Artificial intelligence can expand the practical frontier of capture, connection, retrieval and presentation, yet most AI-assisted workflows continue to treat a document as the primary knowledge container and a model’s private context as the mechanism of continuity. This paper proposes the semantic artifact: an identified, governed, versioned and provenance-preserving representation of the captured knowledge generated through a bounded body of work. Source records, observations, factual assertions, interpretations, decisions, disagreements and explicit gaps remain distinguishable; original multimodal expression can coexist with structured descriptions; current semantic state is paired with non-erasing history; and papers, reports, lessons, exhibitions and interfaces become traceable projections, while code, plans and other consequential outputs become traceable derivatives requiring additional validation. This arrangement is intended to support epistemic reconstruction: later people or systems can interrogate more of the represented path beneath a final account without claiming access to uncaptured context or subjective experience. Independently governed artifacts may also be federated without silently merging identity or authority. The Semantic Artifact Framework specifies these obligations without prescribing one ontology, storage system, model provider or serialisation. At its broadest, it proposes semantic artifacts as a candidate new knowledge medium rather than merely a documentation technique. The proposal is demonstrated by representing this work first as a semantic artifact and generating the present paper from a declared snapshot. This self-demonstration establishes neither completeness nor superiority; it exposes a research agenda spanning capture and projection fidelity, cross-system interoperability, governance, privacy, domain validation and comparative performance.