• qa_deck_100#slide 42 · Retrieval, tools, adapters, and alignment evidence

In RAIDT, a retrieval query must be captured in the run-level evidence pack because retrieval-augmented generation changes what the model can say at run time. The papers argue that governance fails if an organisation cannot later reconstruct one configured use in context. Evidence Review states that reconstructing a run requires retrieval queries, retrieved sources, and the state of retrieval indexes. Foundations then makes this operational by requiring query or corpus identifiers for retrieval augmentation and a preserved snapshot of retrieved passages with document identifiers and hashes. If the query is absent, reviewers may know that retrieval occurred, but they cannot show what the system actually asked the source space to return.

That matters directly to the five pillars (Responsibility, Auditability, Interpretability, Dependability, Traceability), especially Auditability and Traceability. RAIDT treats the run as the unit of governance, so a later reviewer must be able to inspect how grounding was produced in that specific run, not infer it from model cards or general system documentation. The retrieval query records the retrieval intent and scope: what issue was searched, how the system framed that search, and whether the search was appropriate for the task. It therefore helps convert RAG from a black-box convenience into evidence that can be reviewed, challenged, and scored. In practical RAIDT terms, missing query evidence usually prevents a run from reaching anchors 1=missing / 3=partial / 5=audit-ready, because the score profile cannot show that grounded content was retrieved in a reviewable way. This is why RAIDT treats retrieval evidence as one of the influence methods as governance interventions rather than as optional back-end telemetry.

Consider a public-service eligibility assistant used by a caseworker to draft advice for a claimant. The model cites policy text in its answer, but the organisation has not stored the retrieval query. Months later, the advice is challenged. Reviewers can see the output and perhaps the cited paragraph, but they cannot determine whether the system searched the right policy corpus, whether the query omitted a relevant condition, or whether the returned passage was only one of several competing provisions.

If the run-level evidence pack contains the retrieval query, the case becomes reviewable. Reviewers can compare the recorded search against the staff member's task, inspect whether the search terms were too narrow or too broad, and judge whether the output was grounded appropriately. That is the difference between a plausible answer and an auditable one.

• 08-RAIDT_Foundations_M_V50
• 13-RAIDT-Evidence-Review_M_v10