• qa_deck_100#slide 18 · Why current governance leaves a run-level gap

Current AI governance leaves a run-level gap because most governance artefacts still sit at the level of principles, models, systems, or periodic organisational process. The papers argue that these layers matter, but they are too coarse to explain what happened in one concrete use event. In generative AI, risk materialises at run time: the same base model can behave differently because of prompt wording, template version, decoding settings, retrieved passages, tool calls, adapters, alignment controls, and the pattern of human oversight. When those run-time determinants are not preserved, governance remains descriptive rather than reconstructable.

RAIDT names this as a residual evidence gap. The problem is not only whether an organisation has a policy, a model card, or an audit routine, but whether it can later show what happened in a specific run and justify reliance on that output. Without a run-level evidence pack, contested uses cannot be properly reconstructed, compared, or challenged. That is why RAIDT treats the run as the unit of governance and uses the five pillars (Responsibility, Auditability, Interpretability, Dependability, Traceability) to make governance readiness observable in a score profile. The papers are especially clear that influence methods as governance interventions intensify this gap: retrieval augmentation, PEFT or LoRA, structured prompting, and preference-based alignment all change behaviour, yet many existing governance approaches do not require those changes to be logged as auditable run artefacts.

A public-service officer uses GenAI to interpret an eligibility rule for a claimant. The department may already have an AI policy, a model procurement record, and an annual audit plan, yet none of those artefacts shows which exact clause was retrieved on that day, which prompt template framed the question, or which model version generated the advice. If the claimant later challenges the decision, the organisation may be unable to demonstrate what information actually shaped the answer.

Under RAIDT, the same event would be preserved as a run-level evidence pack containing the prompt, model and tool configuration, retrieval snapshot, output, and review steps. The officer's run could then be scored across the five pillars, making it possible to see whether the advice was merely plausible or genuinely audit-ready and contestable.

• 08-RAIDT_Foundations_M_V50
• 11-RAIDT_Academic_Logic_M_v11