• qa_deck_100#slide 19 · Why current governance leaves a run-level gap

GenAI makes the governance problem sharper because it has moved into routine organisational work where outputs can shape records, recommendations, decisions and stakeholder communications. Across the papers, the issue is not simply that organisations use AI more often; it is that GenAI behaviour is materially shaped at run time. A formally approved model can still behave differently across uses because prompts, retrieved context, enabled tools, adapters, safety settings and reviewer actions vary from one case to another. Governance therefore becomes a problem of reconstructing one configured use, not merely declaring that policy, model documentation or lifecycle controls exist.

The problem is sharper still in uncertain and high-stakes settings. The managerial paper argues that organisational users often work with incomplete, ambiguous or misleading information, while GenAI can produce fluent outputs that look authoritative even when grounding is weak. In such contexts, explainability alone is insufficient; uncertainty communication, human oversight and visible checks are also needed. The governance papers therefore argue that higher-level controls remain necessary but are too coarse on their own, because they describe what should happen rather than what can be shown to have happened in one contested case.

RAIDT addresses this by treating the run as the unit of governance. Each material use should yield a run-level evidence pack and a score profile across the five pillars (Responsibility, Auditability, Interpretability, Dependability, Traceability), using anchors 1=missing / 3=partial / 5=audit-ready. This is also why RAIDT treats influence methods as governance interventions: prompting, retrieval, LoRA/PEFT and alignment do not merely change performance, but alter what can be inspected, justified, repeated and challenged.

A public-service team uses GenAI to draft housing-benefit eligibility advice for caseworkers. The same base model may give different advice depending on the prompt template, the exact rule text retrieved, and whether a reviewer edits or approves the answer before it reaches the applicant. If the applicant later challenges the advice, an AI policy or a model card will not be enough.

The organisation needs the run-level evidence pack for that specific case: prompt version, retrieval snapshot identifiers and hashes for the rules used, active settings, the generated answer, any uncertainty statement, and the reviewer?s approval or override note. In RAIDT terms, governance becomes sharper because a plausible answer without this evidence cannot score well on Auditability or Traceability, even if it sounded confident at the time.

• 01-Responsible_AI_for_Managerial_Decision-Making_Under_Uncertainty-V3
• 11-RAIDT_Academic_Logic_M_v11
• 15-RAIDT-IS-Governance_M_v07