• qa_deck_100#slide 81 · Implementation modes and deployment choices

When RAIDT is implemented through automated orchestration, the main shift is from retrospective, partly narrative governance to workflow-level evidence capture around the run as the unit of governance. The papers describe RAIDT as a run-level method that produces a run-level evidence pack and a score profile across the five pillars (Responsibility, Auditability, Interpretability, Dependability, Traceability). Under manual implementation, reviewers may populate records after the event. Under automated orchestration, wrappers or orchestration layers record run identifiers, prompt versions, model and tool settings, retrieval snapshots, hashes, outputs, checks, and review states as the run occurs. This makes governance readiness inspectable from logged artefacts rather than from free-text recollection.

What therefore changes is not the conceptual logic of RAIDT, but its operational reliability, comparability, and scalability. Evidence completeness can be checked automatically, provisional scoring can be suggested for Auditability and Traceability, and repeat-run stability testing can support Dependability. In the empirical paper, influence methods as governance interventions are shown to alter what can be reconstructed and contested; orchestration makes those differences visible at scale because the evidence is captured systematically rather than ad hoc. The result is governance-by-construction: low-scoring runs can trigger instrumentation fixes, escalation rules, or workflow redesign. Read through the anchors 1=missing / 3=partial / 5=audit-ready, orchestration moves organisations away from missing or fragile evidence and towards consistently reviewable, audit-ready run records, while still leaving judgement-heavy elements to calibrated human review.

In a public-service eligibility chatbot, a manual process might leave staff with only the final answer and a vague memory that retrieval was used. Automated orchestration changes that. The wrapper logs the run ID, prompt template version, model deployment, retrieval query, the exact retrieved policy passages, their identifiers and hashes, the output text, and the reviewer decision. RAIDT can then generate a score profile showing whether the advice is auditable and traceable enough for contested use.

If a claimant later challenges the advice, the organisation can reconstruct the exact policy text that informed the response rather than relying on a general citation. If scores fall because retrieval snapshots or review checks are missing, the issue is visible as a governance defect in the workflow itself, not merely as an isolated staff failure.

• 08-RAIDT_Foundations_M_V50
• 09-RAIDT_Empirical_M_V50.docx