For an LLM, Origami changes the document from something that must largely be reverse-engineered from reflowed prose into arbitrary chunks into something that can be interrogated as a structured, addressable knowledge object. The central benefit is not simply that an AI can read the text; it is that the AI can ground its interpretation and output in source-defined structure and return evidence that a person can inspect. • Retrieve stable semantic elements rather than arbitrary token chunks. • Use canonical element addresses as durable retrieval keys. • Keep retrieval results connected to the exact source passage, figure, equation, table, note, or definition. • Cite exact source elements in generated answers. • Make AI citations directly navigable in an Origami-aware reader. • Check quotations against the addressed source element rather than trusting a generated quote. • Use author-supplied concept definitions before inventing or inferring meanings. • Distinguish a defined concept from a coincidental keyword occurrence. • Use canonical bibliography records rather than reconstructing references from formatted text. • Follow explicit citation and cross-document relationships. • Use document structure and heading indexes instead of inferring hierarchy from typography alone. • Read table formulas separately from published table values. • Inspect structured MathML rather than treating equations as images or unstructured strings. • Inspect 3D figure metadata, source relationships, scale, orientation, and other authored facts. • Distinguish authorial semantics and authored layouts from reader-created annotations and AI-generated interpretations. • Avoid contaminating the source publication with agent-generated notes because AI annotations can remain external overlays. • Compare editions using work identity, stable element identity, and explicit lineage evidence. • Distinguish an unchanged element from a replacement or an inferred successor. • Build RAG indexes around persistent document identities instead of temporary chunk numbers. • Reduce brittle re-chunking when typography, layout, or pagination changes. • Preserve provenance through retrieval, synthesis, and answer generation. • Create summaries whose claims point back to exact supporting source elements. • Create cross-document syntheses with element-level source trails. • Return an evidence chain for questions such as “Why do you say that?” or “Show me the evidence for this claim.” • Build semantic retrieval over concepts, citations, people, equations, figures, and relationships rather than lexical similarity alone. • Use the visible Visual-Meta colophon as a discovery fallback when an ingestion pipeline initially receives only extracted text. • Escalate from plain text to richer records: read text, discover that structured metadata exists, retrieve the declared records, then answer from precise elements. • Operate on the publication without proprietary application APIs. • Support local or on-device models over the same structured publication. • Build temporary AI-generated maps, comparisons, study guides, or analytical views while clearly distinguishing them from authored structures. • Make AI output auditable against the publication rather than merely plausible. • Provide a stronger basis for agentic document workflows because an agent can discover the profile, records, authority rules, addresses, and relationships programmatically.