A real accession beside substituted prose produces an apparently real citation. Every identity field genuine, the text from somewhere else. Nothing about it looks wrong. So identity and content are bound together before content can be classified, and a classifier never receives loose text alongside separate metadata.
The document is fetched from the named source and its bytes hashed at the moment of retrieval. A document that carries an accession must also carry the canonical URL it came from.
Source, accession, URL and content hash are combined into one document identity. A document cannot be constructed with a real accession and unverified text at the same time.
SEC_EDGAR · VERIFIED
The cited passage is found in the retrieved text at a recorded offset, rather than trusted. A model can fabricate a quotation as easily as a conclusion, and a fabricated span is worse than none because it looks like evidence.
Interpretation happens last and stays attached to the span that produced it. Synthetic content is representable: demos need it, but it can never inherit an accession, a URL, or a verified badge.