• qa_deck_100#slide 99 · Empirical programme, calibration, procurement, and assurance

In RAIDT, assurance means making governance claims inspectable at the level where generative-AI use actually becomes consequential. Rather than accepting policy statements, model cards, or supplier assurances as sufficient, RAIDT treats the run as the unit of governance and asks whether one configured use can be reconstructed, reviewed, challenged, and compared. The central assurance artefacts are the run-level evidence pack and the score profile across the five pillars (Responsibility, Auditability, Interpretability, Dependability, Traceability). In that sense, assurance is not a vague claim that a system is 'responsible'; it is a disciplined way of showing, with preserved evidence, what happened in a particular run and whether the governance conditions for that run were strong enough to justify use.

This makes assurance run-specific. A proper run-level evidence pack records the prompt or instruction, model and tool configuration, retrieved context where relevant, output, output hash, timestamps, and the human or automated checks applied. The score profile then interprets that evidence using anchors 1=missing / 3=partial / 5=audit-ready, so governance readiness becomes observable rather than merely asserted. RAIDT therefore positions assurance as evidence-based review, not as a guarantee of factual correctness or legal compliance in itself. Its purpose is to make runs reviewable, contestable, and improvable over time, while keeping trade-offs visible across the five pillars rather than hiding them inside a single unexamined compliance judgement.

In a healthcare setting, a hospital may use a generative system to support black-box clinical decision support or to draft a clinician-facing explanation. If a recommendation is later challenged, RAIDT assurance does not begin with the supplier's general documentation. It begins with the specific run-level evidence pack for that use: the exact prompt, model version, retrieved material, output, and the checks completed before the result was relied upon.

The hospital can then inspect the score profile across the five pillars (Responsibility, Auditability, Interpretability, Dependability, Traceability). If Auditability is low because the retrieval snapshot was not preserved, or Responsibility is weak because no appropriate reviewer signed off, the governance gap is visible immediately. Assurance therefore means that the disputed run can be examined as evidence, not reconstructed from memory or defended only by broad policy claims.

• 10-RAIDT_Policy_Pathways_M_V50
• 16-RAIDT-Audit-Accountability_M_v05