• workshop_dense_100#slide 8

RAIDT is needed because most responsible-AI approaches for GenAI remain strong on principles but weak on reconstructable proof. The papers argue that organisations often rely on policy statements, model cards, explainability artefacts, or episodic audits, yet these are rarely sufficient to show what happened in one material use event. For GenAI, that weakness is acute because outputs are shaped at run time by prompts, templates, retrieved context, tools, model settings, and oversight decisions. When those artefacts are not preserved, a challenged output may be fluent but not governable: the organisation cannot reliably reconstruct the event, test whether controls were active, or show which evidence informed the response.

RAIDT addresses this residual run-level evidence gap by establishing the run as the unit of governance and requiring a run-level evidence pack as the standard proof object. That pack captures the configured use in context, including prompt and template versions, model and tool configuration, retrieved passages where applicable, outputs, checks, and oversight. The five pillars (Responsibility, Auditability, Interpretability, Dependability, Traceability) then convert that evidence into a score profile, using anchors 1=missing / 3=partial / 5=audit-ready. This makes governance readiness inspectable and comparable across runs rather than asserted narratively.

RAIDT is also needed because contemporary governance increasingly depends on evidence that can be reviewed after the fact. The papers therefore frame influence methods as governance interventions: retrieval augmentation, structured prompting, PEFT/LoRA, and preference-based alignment can improve outcomes, but only if they are logged and reviewable. In that sense, RAIDT connects principles to operational accountability, supports contestability and audit sampling, and enables organisations to learn from repeated use rather than merely declaring that controls exist.

In a public-service eligibility workflow, a caseworker uses GenAI to interpret benefit rules for an applicant with mixed employment history. Without RAIDT, the final advice may look confident, yet the organisation may not know which policy clause was retrieved, which prompt template framed the question, or whether the model version changed between cases. If the applicant challenges the outcome months later, the team has little more than a narrative explanation.

With RAIDT, the same interaction is stored as a run-level evidence pack containing the run ID, prompt template, model version, retrieval query, exact rule passages with document identifiers and hashes, output, and reviewer sign-off. The score profile can then show whether Auditability and Traceability remain partial or are closer to audit-ready for that specific run. This matters because the organisation can reconstruct the reasoning context, verify the policy version used, and improve logging or oversight where the evidence remains weak.

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