There's a category error in many LLM systems: retrieving documents, putting them in context, and calling the answer "grounded." Retrieval gives access to candidates for evidence, but it doesn't prove that a retrieved passage supports a claim. A passage can be stale, incomplete, topically similar but non-entailing, conflicting, untrusted, or maliciously injected. The original RAG paper identified provenance as an open problem.

The post-RAG objective must be an evidence-grounded narrative system where every material proposition gets inspectable support, every business decision is traceable, and uncertainty has a first-class representation. Optimizing the evidence system creates an artifact a developer can inspect. The new unit of control is the atomic claim, not documents. A claim ledger lists all atomic claims with evidence and controls: a stable claim_id, claim text with exact span, verbatim evidence spans, canonical source with content hash, retrieval context, and a support decision (directly supported, permitted synthesis, insufficient evidence, conflicting evidence, unsupported).

The non-negotiable publication gate: a material claim without acceptable support is a request to revise, abstain, label as inference, or escalate. This forces a product decision between unsupported fluency and visible uncertainty.