• workshop_dense_100#slide 29

In RAIDT, the literature gap is the residual run-level evidence gap: the mismatch between mature responsible AI principles, audits, documentation, and observability on one hand, and the absence of a standard, inspectable proof object for one configured use on the other. Existing literature can say what good governance should look like, or record fragments of technical activity, yet still leave an organisation unable to show exactly what happened in a disputed run. RAIDT treats this as a real gap in governance granularity rather than a minor implementation detail. Generative AI outcomes are materially shaped at run time by prompt versions, retrieval settings, model versions, tool access, adapters, and human oversight.

A simple example is a credit adverse-action explanation generated for a customer. A firm may have an approved model card, an internal audit programme, and a policy requiring explainability. Even so, if it cannot later show the structured prompt version, the model deployment ID, the reason-code template, the retrieved policy or data snapshot, the output hash, and the reviewer check, the explanation cannot be reconstructed or contested properly. RAIDT therefore defines the run as the unit of governance and uses a run-level evidence pack plus a score profile across the five pillars (Responsibility, Auditability, Interpretability, Dependability, Traceability). This matters because governance claims become evidence-based rather than rhetorical: anchors 1=missing / 3=partial / 5=audit-ready show whether the organisation can actually support review, comparison, escalation, and learning.

A bank uses GenAI to draft adverse-action explanations after a lending decision. The system is technically capable and the bank has enterprise AI policies, but a customer disputes the explanation and asks why a particular reason was given. If the bank only retains the final text, it cannot show which prompt schema was used, whether a retrieval step inserted an outdated policy fragment, or whether a human reviewer amended the draft before sending it.

RAIDT matters here because it reframes the gap as an evidence problem, not just a documentation problem. A complete run-level evidence pack would let compliance, audit, and business teams reconstruct the event, inspect the evidence, and assign a defensible score profile instead of relying on a general statement that the system was governed.

• 08-RAIDT_Foundations_M_V50
• 16-RAIDT-Audit-Accountability_M_v05