The AI-ready deliverable is a RAG-ready corpus AND a knowledge graph, both as portable files

Export produces two coherent artifacts on Storage, never a live database. (1) A RAG-ready corpus: cleaned, chunked, Screened, PII-handled text carrying full Provenance and Classification, written as Parquet/JSONL; embeddings are optional (an opt-in Embedder Provider) so any vector DB can ingest it. (2) The knowledge graph — nodes, edges, and per-edge Evidence — as portable files (Parquet + TTL + GraphML) that load into Neo4j, a triple store, or a graph-RAG engine. Serving both vector-RAG and graph-RAG consumers without picking their stack is deliberate: leading KG-only narrows the buyer (most enterprise AI spend is vector-RAG today), while corpus-only discards the corpus-level Disambiguation + KG that is the differentiator versus unstructured/docling. Writing directly into an adopter's Neo4j/vector DB/triple store is rejected for v1 — it would make the framework own DB connections, schemas, and upserts, the exact vendor-coupling and ops surface ADR-0003 and ADR-0009 avoid; such writers are opt-in Provider work post-v1.