• workshop_dense_100#slide 14

RAIDT may be defined as a run-level evidence framework for responsible governance of generative AI in organisational work, designed to turn a configured use into a reconstructable and scoreable governance object. In substantive terms, it requires a run-level evidence pack containing the artefacts needed to reconstruct what happened, and it expresses governance readiness through a score profile across the five pillars (Responsibility, Auditability, Interpretability, Dependability, Traceability). Because the scored object is the run and not the model in general, RAIDT gives organisations a disciplined way to evaluate how a system was actually used in context.

An example from the papers is public-service eligibility advice. Without RAIDT, staff may keep only the answer, leaving no reliable record of which rule text, prompt version, or model configuration produced it. With RAIDT, the run-level evidence pack stores the prompt, model version, retrieval snapshot identifiers and hashes, output, and checks, after which the run receives a score profile. This matters in RAIDT because the framework's whole value lies in making governance inspectable at the point where risk materialises. The papers repeatedly argue that principles, model cards, and episodic audits do not by themselves provide contestable proof of what happened in one use. RAIDT matters because it supplies that proof and links it to a structured judgement using anchors 1=missing / 3=partial / 5=audit-ready.

Imagine a local authority officer using GenAI to answer a citizen's question about benefit eligibility. If the authority keeps only the final advice, it cannot later show which policy version informed the answer or whether retrieval grounding was actually used. That creates a governance weakness even if the text appears sensible.

Under RAIDT, the officer's configured use becomes a governed run. The evidence pack records the retrieval query, the rule passages retrieved, document identifiers and hashes, the prompt, the output, and the review steps. Reviewers can then score auditability and traceability with reference to preserved artefacts rather than memory. The example shows why RAIDT matters: it converts an otherwise fleeting interaction into an accountable organisational record.

• 00-RAIDT_Wording_v2
• 08-RAIDT_Foundations_M_V50