Mapping WYSIWYG Document Diffs to AI Runs and Sources

Summary

Lexifina describes a version-control design for making AI edits to WYSIWYG documents granular and auditable. The article notes that formats such as DOCX combine wording, formatting, and structure in a shared OOXML representation, so a useful diff must distinguish harmless serialization changes from changes visible in the document. The proposed system links each change to recorded actions including an AI proposal, a reviewer’s decision, and the final application, while preserving full agent traces. It snapshots the document state, agent run and supplied context, proposal details, review and application records, and later history. These records can show what an agent could read, what it intended to change, what entered the saved document, and which contribution explains the current wording. The article covers text, formatting-only, image asset, image presentation, structural, notes and links, and serialization-only diffs, with comparison methods suited to each type. It also argues that tracking added or modified paragraphs and rejected proposals can support persistent graph, vector, or metadata representations for information retrieval.