• qa_deck_100#slide 80 · Implementation modes and deployment choices

Semi-automated RAIDT is the right implementation mode when an organisation has moved beyond an exploratory pilot, but is not yet operating a fully instrumented orchestration layer. Across the papers, RAIDT works with run as the unit of governance and requires each material use to yield a run-level evidence pack and a score profile across the five pillars (Responsibility, Auditability, Interpretability, Dependability, Traceability). Semi-automation is therefore appropriate where the organisation can reliably log core run metadata and evidence fields, yet still needs human review for the more judgement-heavy parts of scoring. This is especially suitable when outputs influence decisions, records, communications, or services, and when the workflow is repeated often enough that purely manual evidence capture becomes inefficient or inconsistent.

The papers also imply a risk-and-maturity threshold. Semi-automated RAIDT is useful when objective checks can be mechanised, such as verifying run IDs, prompt versions, model deployment identifiers, retrieval snapshot pointers, hashes, and recorded checks, but where responsibility, interpretability, and dependability still require contextual judgement. In that sense, semi-automation is a proportionate implementation mode: it reduces the burden of evidence handling without pretending that governance can be delegated to a model. It is particularly well suited to settings where reviewability must be improved quickly, where multiple stakeholders may later inspect the run, and where evidence capture can be performed lawfully under access controls. The papers are also clear that implementation mode does not alter the governance object: the evidence pack remains the auditable record, and the substantive standard is still whether the run can be reconstructed, reviewed, and justified against the anchors 1=missing / 3=partial / 5=audit-ready.

In a healthcare note-summarisation workflow, a hospital may already have a wrapper that records the prompt template ID, model deployment ID, decoding settings, output hash, retrieval snapshot hash, and whether a nurse reviewed the draft. That is enough infrastructure to support semi-automated RAIDT, even if the organisation is not ready for a full evidence repository integrated across all services. The system can assemble the metadata automatically, while a clinician or governance reviewer checks whether uncertainty was communicated, whether escalation was documented, and whether the summary is suitable for triage.

This is the right mode because the workflow is repeated, high-impact, and review-sensitive. Manual capture alone would be burdensome and prone to omission, while full automation would be premature if the hospital still wants human judgement over safety-relevant reliance decisions.

• 08-RAIDT_Foundations_M_V50
• 09-RAIDT_Empirical_M_V50.docx