• workshop_table17#tag-band C0 · 0–20 min

RAIDT is a run-level evidence framework for responsible governance of generative AI in organisational work. Its basic claim is that governance should attach to the material use event, not only to the model, the policy, or the organisation's general assurance narrative. Accordingly, RAIDT treats the run as the unit of governance and operationalises governance through a run-level evidence pack and a score profile. The run-level evidence pack captures the bounded artefacts needed to reconstruct and review what happened in a particular use, while the score profile expresses governance readiness across the five pillars (Responsibility, Auditability, Interpretability, Dependability, Traceability).

Its novelty lies in combining those two elements into a single governance method: a bounded proof object plus an anchored governance-readiness measure using anchors 1=missing / 3=partial / 5=audit-ready. This allows governance to be inspected, compared and improved across runs, configurations, suppliers and domains. It also means that influence methods as governance interventions are assessed as part of governance, because prompting, retrieval, adaptation and alignment shape both behaviour and evidential quality. RAIDT is therefore evidence-based rather than vendor-based, and measurement-based rather than purely declarative.

RAIDT is not a single software product, and it is not merely a request for a model to write an evidence summary. It is not a general theory of all trust in AI, not a guarantee that an output is correct, and not a replacement for domain safety practice or formal decision procedures. In healthcare it does not replace clinical judgement; in finance it does not replace legal duties; in public services it does not replace formal administrative process. Its role is narrower and more practical: to make governed GenAI use reconstructable, reviewable and improvable.

In healthcare note summarisation, a clinician may use GenAI to draft a structured summary for a patient presenting with chest pain. RAIDT does not act as the clinician, and it does not certify that the summary is medically correct. Rather, it requires a run-level evidence pack showing the prompt constraints, any uncertainty requirement, the model and tool configuration, and whether escalation or human oversight was specified.

That run can then receive a score profile across the five pillars (Responsibility, Auditability, Interpretability, Dependability, Traceability). If preference-based alignment was used to encourage conservative behaviour, that matters because influence methods as governance interventions must be logged. The result is a governable record of one clinical support event. The clinician still exercises judgement, but the organisation can later inspect whether the run was bounded, reviewable and audit-ready, rather than relying only on a fluent summary or a general policy statement.

• 00-RAIDT_Wording_v2
• 08-RAIDT_Foundations_M_V50
• 11-RAIDT_Academic_Logic_M_v11