Semantic Persistence for LLM-Mediated Workflows Proposed.

A conceptual model for Large Language Model (LLM)-mediated workflows has been introduced, emphasizing "semantic persistence." This model posits that not only the outputs but also the workflow definitions, their instances, inference records, context snapshots, and dependency relations should be represented as inspectable, resumable, and reviewable knowledge objects within a shared knowledge substrate. Drawing inspiration from Lisp's symbolic forms and live-image thinking, the model aims to provide a language-independent framework for managing the complexity of LLM applications that involve tool use, retrieval, branching logic, checkpointing, and human approval. It moves beyond traditional workflow systems by treating the workflow itself as a form of persistent knowledge. A central distinction in this model is made between "derive" and "infer." "Derive" refers to deterministic computations based on available state, while "infer" denotes LLM-mediated judgments made under declared context and executor-controlled capability policies. This preliminary account lays the groundwork for future work on formal transition semantics, offering a new perspective on building more robust and transparent LLM applications.