• qa_deck_100#slide 73 · RAG, PEFT, RLHF-type controls, and stacked configuration

RAG is a governance intervention in RAIDT because it changes the evidential structure of a system run, not merely the information available to the model. In RAIDT, influence methods as governance interventions are assessed against the five pillars (Responsibility, Auditability, Interpretability, Dependability, Traceability), and the central question is whether a method leaves inspectable artefacts that can support oversight, contestation, and replay. The RAG paper argues that provenance-first RAG externalises knowledge into citeable, inspectable artefacts at inference time: retrieved documents, chunk identifiers, source URIs, retrieval scores, corpus snapshots, prompt versions, and output hashes. That materially strengthens Interpretability and Traceability, and it can also raise Auditability because the output is no longer a free-floating answer detached from evidence lineage.

This is why RAIDT treats RAG as more than a retrieval feature. The practical focus is the run as the unit of governance: each run should emit a run-level evidence pack that makes the result reviewable after the fact. In the provenance-first design, that pack includes the prompt lineage, retriever settings, retrieval context IDs, corpus or index snapshot, citations attached to claims, timestamps, reviewer scores, and SHA-256 hashes. On a RAIDT score profile, using anchors 1=missing / 3=partial / 5=audit-ready, prompt-only outputs may still read well yet remain weakly contestable. RAG moves the system towards audit-ready status because reviewers can inspect where claims came from, test whether the cited material really supports them, and reconstruct the exact conditions of the run. The papers also stress that this governance value depends on disciplined logging and index governance; stale or poor-quality retrieval can degrade the very controls RAG is meant to provide.

In a clinical decision-support setting, a hospital uses an LLM to turn de-identified consultation notes into four bullets on symptoms, diagnosis, treatment, and red flags. Without RAG, a clinician may receive a fluent summary but cannot tell whether a warning came from the note, from an approved guideline, or from the model's own unsupported completion. The run-level evidence pack is therefore thin, and disputes are difficult to resolve.

With provenance-first RAG, the system retrieves note spans and approved clinical guidance, inserts content IDs into the summary, logs the FAISS snapshot, top-k document IDs, similarity scores, prompt version, and output hash, and stores reviewer scores. If a sepsis warning is missed or miscited, the team can reconstruct the exact run, inspect the retrieved passages, and decide whether the problem arose in retrieval, prompting, or reviewer judgement. That is governance work, not merely better search.

• 06-RAIDT_RAG_V1
• 07-RAIDT_RLHF_V1