Provenance for RAG: citations your users can verify

Retrieval-augmented generation fetches documents and hands them to a model. The model writes an answer. The common failure is not retrieval, it is accountability: the answer reads as confident, the sources sit in a vector store, and a user who wants to check a claim has no path from the sentence to the source.

Provenance closes that path. Three additions turn a RAG answer into one a reader can verify.

Inline citations that resolve

Every claim in the answer points to the retrieved passage it came from, and the pointer opens the actual source, not a reranked snippet. A citation that cannot be opened is decoration.

A record of what was retrieved, and when

The set of documents the model read is itself a fact worth keeping. Retrieval changes as the index changes; a receipt over the retrieved set lets you reproduce the answer or explain why it moved.

A signature over the answer and its sources

Binding the output to the exact inputs means a reviewer can confirm the answer was generated from those documents and not edited after. It does not make the answer correct; it makes it accountable.

The payoff shows up in the hard conversations. When a user disputes an answer, you can show the passages it stood on and the time they were read. When a regulator asks how a system reached a conclusion, you have the chain instead of a shrug. None of this requires a new model. It is plumbing around the one you have: cite to the source, record the retrieval, sign the result.

DRM3 runs this pattern as a product. Ask the record a question and the answer comes back with its sources cited inline and a receipt over the answer: trusted data for AI, with citations your users can check.