• qa_deck_100#slide 43 · Retrieval, tools, adapters, and alignment evidence

A retrieval snapshot or index ID proves which source space, and which state of that source space, was actually available to the system when the run occurred. The papers are clear that citations alone do not achieve this. In Foundations, the public-sector eligibility example states that auditability depends on reconstructing the exact rule text used, requiring retrieval snapshot identifiers and hashes rather than citations alone. Evidence Review adds that reconstructing a run with retrieval requires not only retrieved sources but also the state of retrieval indexes. Technical Foundation similarly argues that traceability depends on preserved retrieval snapshots and identifiers because retrieved documents can change or disappear over time.

Accordingly, a retrieval snapshot or index ID proves more than the fact that RAG was enabled. It fixes the provenance boundary for that run: which corpus build, which retrieval state, or which preserved passage set grounded the output. When linked to document IDs, timestamps, and hashes, it allows later reviewers to verify whether the cited material was genuinely available at the time, whether the corpus later drifted, and whether the retrieved evidence remained intact. This is why RAIDT includes such identifiers in the run-level evidence pack and score profile. Within the five pillars (Responsibility, Auditability, Interpretability, Dependability, Traceability), the immediate effect is strongest for Auditability and Traceability, with knock-on value for Dependability because repeated runs can be compared against stable retrieval states. Without snapshot or index evidence, reviewers are left with a narrative claim about grounding rather than proof, which is inconsistent with anchors 1=missing / 3=partial / 5=audit-ready.

Imagine a cybersecurity alert triage assistant that retrieves internal playbooks and threat notes before recommending next steps. The organisation stores the output and the cited note title, but not the index ID. Two months later the threat corpus has been refreshed, several notes have been merged, and one indicator entry has been corrected. The team now cannot prove whether the assistant acted on the old note, the corrected note, or a different index build altogether.

If the run-level evidence pack includes the retrieval snapshot or index ID, reviewers can bind the recommendation to the exact corpus state used at the time. They can then inspect the retrieved documents and determine whether the recommendation was grounded in the right evidence, even though the live system has moved on.

• 08-RAIDT_Foundations_M_V50
• 13-RAIDT-Evidence-Review_M_v10
• 18-RAIDT-Technical-Foundation_M_v04