• qa_deck_100#slide 79 · Implementation modes and deployment choices

Manual RAIDT implementation is a low-tech but disciplined operating arrangement for evidence-centred governance. The papers argue that RAIDT treats the run as the unit of governance, so even without orchestration tooling the organisation still assembles a run-level evidence pack for each material use. In practice, that means preserving the prompt or template version, the output, relevant configuration notes, any retrieved context, reviewer observations, and oversight decisions in a form that can later be reconstructed and inspected. In early pilots, the papers explicitly state that teams may populate run records manually and apply the rubric by hand; the item definition translates this into controlled spreadsheets, saved prompts/outputs, reviewer forms and scoring sheets.

Manual implementation is therefore not informal governance. A reviewer still scores the five pillars (Responsibility, Auditability, Interpretability, Dependability, Traceability) and records evidence pointers that justify the judgement. The resulting score profile should be grounded in the anchors 1=missing / 3=partial / 5=audit-ready, so the assessment remains comparable across runs. Crucially, RAIDT also requires influence methods as governance interventions to be documented: prompting, retrieval augmentation, adaptation layers, and alignment controls must be recorded because they shape what happened in the run.

This matters because manual implementation gives organisations a credible starting point before automation. It supports sampling, calibration, challenge, and remediation, while revealing which evidence fields are repeatedly absent and therefore need stronger instrumentation later.

In a public-service eligibility advice pilot, a team uses a controlled spreadsheet as the run register. For one claimant interaction, the assessor saves the prompt used, the model output, the policy extract consulted, the retrieval snapshot identifier if applicable, and a short reviewer note explaining whether the output cites the correct clause version. A second reviewer checks whether the saved material is sufficient to reconstruct the run and assigns pillar scores on a scoring sheet.

If the case shows citations but no preserved policy snapshot, Auditability and Traceability cannot be treated as strong. The manual process makes that gap visible immediately. The team can still produce a defensible score profile, escalate the run for supervisory review, and update its evidence-capture routine before wider deployment.

• 08-RAIDT_Foundations_M_V50
• 18-RAIDT-Technical-Foundation_M_v04