nt model stores canonical content plus an explicit structure tree plus in-band provenance, so machines need no layout extraction. JATS ANSI NISO Z39.96-2015 is a proven structure-native scholarly format whose markup is structural semantic and declarative not presentational. JATS carries provenance in-band: DOI, ORCID, ROR ISNI Ringgold affiliation identifiers, CRediT contributor roles, and structured author names. IPLD is the content-addressed substrate: a CID is a multihash plus a multicodec, blocks self-describe their codec and link by hash. IPLD does not guarantee byte-deterministic content identifiers for the same logical object, so a content-addressed store must canonicalize bytes before hashing. Content canonicalization for dedup needs strict DAG-CBOR or sorted keys or normalized bytes plus stable sub-span byte-range addressing so a fact can cite a byte range. Presentation formats like PDF and EPUB bury structure in layout and are lossy for machine retrieval; VILA shows recovering layout structure improves token classification by 1.9 percent Macro F1. Store document structure natively and treat PDF EPUB and HTML as lossless export targets for compatibility, not as the storage format.