• integrated_82#Q4.11

RAIDT can be used in law, public services, and social-care-style settings by treating the run as the unit of governance rather than treating assurance as a generic policy statement. For each significant use, the organisation captures a run-level evidence pack covering prompt or template identifiers, model and tool settings, retrieved context where relevant, outputs, and recorded checks. That run is then assessed through the five pillars (Responsibility, Auditability, Interpretability, Dependability, Traceability), producing a score profile that makes governance readiness inspectable at the point of use. This is especially valuable where outputs may shape case handling, eligibility explanations, safeguarding notes, or legally sensitive drafting.

In the papers, these sectors are presented as settings where accountability, procedural fairness, review, and contestability matter more than fluent text alone. RAIDT therefore treats influence methods as governance interventions: structured prompting can improve understandable reason-giving, RAG can improve auditability and traceability when retrieval snapshots are retained, and LoRA or stacked configurations can improve stability only if their versions and checks are logged. In ageing and public-interest services, this is further calibrated around vulnerability, inclusion, and the ability to challenge decisions, so users receive reasons, policy bases, escalation routes, and timely correction rather than opaque automation. Operationally, teams can sample runs, score them with anchors 1=missing / 3=partial / 5=audit-ready, and use low scores to trigger human review, remediation, or procurement challenge. In legal and public-service work, RAIDT is therefore less a claim that AI is trustworthy in the abstract than a method for making each use reconstructable, reviewable, and challengeable.

A local authority deploys a GenAI assistant to draft an eligibility explanation for an older adult seeking transport and care support. The assistant uses structured prompting plus RAG over the current policy text. For each reply, staff capture a run-level evidence pack containing the run ID, prompt version, retrieved rule passages with dates, model configuration, output text, and reviewer notes. The resulting score profile shows whether the explanation is understandable, whether the cited policy can be reconstructed, and whether uncertainty or escalation to a human adviser was made explicit.

If Auditability or Traceability falls below an acceptable threshold, the message is not sent. Instead, a caseworker revises the explanation, confirms the applicable policy version, and ensures the citizen receives reasons, next steps, and a clear appeal route. This mirrors the papers' public-service and ageing-services logic: RAIDT supports service efficiency, but only where evidence capture, reviewer calibration, and contestability are built into the workflow.

• 09-RAIDT_Empirical_M_V50.docx
• 20-RAIDT_AgeingSoc_M_V50
• 21-RAIDT_Sector_Playbook_Healthcare_V2