• workshop_dense_100#slide 49

In RAIDT, a retrieval snapshot or index ID is a stable identifier for the versioned retrieval space, or for the preserved set of retrieved passages, that grounded a particular run. Its function is temporal and evidential: it fixes the source environment used at the time of generation so that later reviewers can reconstruct what was actually available to the system. Foundations states that auditability in retrieval-based use cases requires retrieval snapshot identifiers and hashes rather than citations alone, and it further recommends recording query or corpus identifiers together with preserved retrieved passages, document identifiers, and hashes. Technical Foundation generalises the same point by arguing that traceability depends on retrieval histories, identifiers, source references, hashes, and retrieval snapshots assembled into a bounded evidence pack.

An example would be a cybersecurity triage assistant that searched a specific indexed build of internal threat guidance during one alert review. The retrieval snapshot or index ID records that exact source state, not merely the fact that some guidance was consulted. This matters in RAIDT because the run-level evidence pack is scored, and the score profile depends on reconstructable provenance. Within the five pillars (Responsibility, Auditability, Interpretability, Dependability, Traceability), snapshot or index evidence is central to Traceability and Auditability, and it also strengthens Dependability by allowing comparisons across repeated runs against stable retrieval conditions. In anchors 1=missing / 3=partial / 5=audit-ready terms, the retrieval snapshot or index ID is what stops grounding evidence from drifting as corpora are updated, corrected, or retired after the original run.

A university support service uses a GenAI assistant to draft student guidance by retrieving internal regulations and departmental procedures. After a complaint, the service needs to show which rule set the assistant relied on when the advice was given. By that point, the regulations site has been updated for the new term and one procedure page has been rewritten.

If the run-level evidence pack includes the retrieval snapshot or index ID, the service can recover the exact retrieval state used in the original run and inspect the retrieved documents linked to it. Reviewers can then determine whether the advice matched the rule set that was live at the time, rather than judging the old output against the current website. That is precisely the kind of reconstructability RAIDT requires.

• 08-RAIDT_Foundations_M_V50
• 18-RAIDT-Technical-Foundation_M_v04