Capstone: Provenance, Audit & Compliance

Output Provenance & Verifiable Citations

Across this track you hardened the inputs to a RAG system — the threat model (Day 1), prompt-injection defense (Day 2), de-identification (Day 3), and the data layer (Day 4). The final lesson is about the outputs: proving, after the fact, exactly how any given answer was produced. In a regulated setting, an answer you can't trace is an answer you can't defend.

What Provenance Means for RAG

Provenance is the record that links a generated answer back to the specific evidence it was built from. For a RAG response, a complete provenance record captures:

• the retrieved chunks that were placed in the context (by stable id), and their relevance scores;
• the source documents those chunks came from (and their version/timestamp);
• the model and prompt template used to generate the answer;
• which guardrails and redactions fired along the way (from Days 2–3).

Citations Are Provenance the User Can See

A citation is the user-facing slice of provenance: "this sentence came from policy-4.2." Citations do double duty — they let a clinician or auditor verify a claim against the source, and they make hallucinations obvious (a confident answer with no supporting chunk is a red flag).

The discipline: never surface a claim the retrieved context doesn't support, and attach the source id to every claim so it can be checked.

Why It's a Compliance Control, Not Just a Nicety

Provenance and audit trails are a recognized expectation for AI in regulated industries. If a system makes a recommendation that influences a decision about a person, you may be required to explain what evidence drove it. Provenance turns "the model said so" into "here is the source paragraph, retrieved at this time, at this relevance score." That is the difference between an auditable system and a liability.