• qa_deck_100#slide 3 · RAIDT core and mindmap orientation

Inside a RAIDT run-level evidence pack sits the minimum recorded proof needed to reconstruct one material use of a GenAI system. Because RAIDT treats the run as the unit of governance, the pack begins with stable identifiers and context: run identifier, timestamp, initiating user role or system role, task label, and domain label. It then records the full prompt history that shaped the run, including system messages, prompt templates, and their version identifiers. The configuration layer is also part of the pack: model identifier, provider, version, decoding settings such as temperature and top-p, enabled tools, tool versions, and any tool outputs that materially informed the final answer.

The pack also contains the evidence needed for review rather than mere description. Where retrieval is used, it stores the retrieval query, index or corpus identifier, and a snapshot of retrieved passages with document identifiers and hashes. It stores the full output, output hashes for integrity checking, the human review steps applied, automated test or safety-filter results, and the retention and access-control metadata that allow secure governance integration. In RAIDT, this is not free-text assurance; it is an auditable record used to score the five pillars (Responsibility, Auditability, Interpretability, Dependability, Traceability). In practice, the pack is what makes the score profile defensible, because the scoring logic rests on anchors 1=missing / 3=partial / 5=audit-ready evidence rather than on retrospective narrative claims.

In a public-service eligibility case, a staff member uses GenAI to interpret a benefits rule. The run-level evidence pack would contain the run ID, timestamp, staff role, the exact prompt and template version, the model version and decoding settings, and the retrieval snapshot of the rule clauses used, with document identifiers and hashes. It would also record the generated advice, the output hash, the reviewer who checked the advice, and the retention tag governing how long the record is kept and who may inspect it.

If a claimant later challenges the advice, the organisation can reconstruct what happened in that specific run rather than relying on memory. That is why the evidence pack sits at the centre of RAIDT: it turns a fluent answer into a reviewable governance object.

• 00-RAIDT_Wording_v2
• 08-RAIDT_Foundations_M_V50
• 11-RAIDT_Academic_Logic_M_v11