• integrated_82#Q1.9

Evidence is central to RAIDT because the framework is built on the premise that organisational governance of GenAI must be demonstrable at the level where risk actually arises: the individual run. In RAIDT, this means treating the run as the unit of governance and asking what can be shown later about a specific use in context, rather than relying on policy statements, model cards, or retrospective assurances. The papers argue that GenAI behaviour is materially shaped at run time by prompts, configuration settings, tool use, retrieved context, and influence methods as governance interventions. For that reason, responsible governance cannot rest on assertion alone; it requires a run-level evidence pack that preserves what happened, under which controls, and with which oversight.

Evidence is also central because RAIDT operationalises governance as measurable outcomes across the five pillars (Responsibility, Auditability, Interpretability, Dependability, Traceability). A score profile is only meaningful if each pillar is justified by inspectable artefacts rather than fluent narrative. This is why RAIDT uses anchors 1=missing / 3=partial / 5=audit-ready: the framework distinguishes weak, incomplete, and reconstruction-ready evidence for later review. In this logic, evidence supports contestability, audit sampling, replayability, and organisational learning, while also reducing the risk of compliance theatre. It allows organisations to compare runs, identify where governance is failing, and improve instrumentation, versioning, and review practice. RAIDT therefore values evidence over assertion because evidence makes accountability reconstructable, comparable, and governable in practice.

A clear public-service example is eligibility advice. A staff member uses GenAI to interpret a benefits rule for a claimant. If only the final answer is kept, the organisation may know what the model said, but it cannot later show which policy clause was retrieved, which prompt template was used, whether the current rule version was applied, or what human review occurred. In that situation, any claim that the advice was responsible remains largely narrative.

Under RAIDT, the same event is captured as a run-level evidence pack containing the prompt, model and tool versions, retrieval snapshot, document identifiers and hashes, output, and review record. The resulting score profile can then be justified against the five pillars (Responsibility, Auditability, Interpretability, Dependability, Traceability). If the retrieval snapshot is missing, Auditability and Traceability may sit nearer anchors 1=missing / 3=partial / 5=audit-ready; if the record is complete, a reviewer can reconstruct the advice and assess whether it was suitable. Evidence is central here because it turns a disputed service interaction into a governable run rather than an unsupported claim about what supposedly happened.

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