• workshop_dense_100#slide 89

Ageing calibration in RAIDT is the sector-specific extension that keeps the five pillars stable while reinterpreting them for ageing-related services through three moderators: vulnerability, inclusion, and ability to challenge decisions. In other words, RAIDT still evaluates the five pillars (Responsibility, Auditability, Interpretability, Dependability, Traceability), and it still keeps the run as the unit of governance, but it asks different threshold questions in health, finance, and social care. A compliant run-level evidence pack must show not only what the model produced, but whether the service interaction was accessible, whether uncertainty and escalation were explicit, and whether a later reviewer or affected person could contest the outcome. The resulting score profile is therefore sensitive to duty of care and procedural fairness, not only to model fluency.

A concrete example appears in the ageing-society healthcare work: baseline prompting generated clinically useful drafts, yet Auditability and Traceability stayed low until instrumentation captured prompt versions, retrieval snapshots, adapter lineage, output hashes, and reviewer logs. The healthcare playbook operationalises this with anchors 1=missing / 3=partial / 5=audit-ready, so RAIDT scoring is attached to observable evidence rather than impressions. This matters because the papers show that influence methods as governance interventions can raise or lower governance readiness depending on what they make reconstructable and reviewable. In ageing services, that distinction is decisive: systems may shape triage, eligibility, or care coordination for people who are more vulnerable, variably digitally included, and more likely to need reasons and routes of challenge.

Consider a bank using GenAI to draft an adverse-action explanation after refusing credit to a pension-age customer. Ageing calibration requires more than a technically correct refusal letter. The explanation must be in plain language, identify the relevant reason codes or policy basis, avoid unnecessary cognitive burden, and show what the customer can do next if they wish to challenge the decision. If staff used structured prompting, the prompt template, model version, decision criteria, generated text, and reviewer notes belong in the run-level evidence pack.

Why this matters is straightforward. Without those artefacts, the bank cannot later show how the explanation was generated or whether the reasons matched recorded criteria. With them, reviewers can inspect the score profile across the five pillars and decide whether the run is suitable for use, requires correction, or should be escalated to a human decision-maker.

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