• qa_deck_100#slide 17 · Why current governance leaves a run-level gap

Responsible AI remains essential, but the RAIDT papers argue that it is still insufficient for governing one GenAI use because most Responsible AI approaches operate at levels that are too general for the evidentiary problem created by generative systems. The literature is strong on principles, lifecycle expectations, model documentation, and periodic audits. It says what an organisation ought to value, such as fairness, transparency, accountability, safety, and oversight. However, it is weaker on what must be captured so that one configured output can later be reconstructed, reviewed, challenged, and compared in context.

That gap matters because GenAI behaviour is materially shaped at run time. A single output may depend on prompt structure, retrieved passages, enabled tools, adapter versions, safety settings, and human review actions. An organisation may therefore have a sound AI policy and still be unable to show what happened in one contested case. In RAIDT terms, higher-level governance remains necessary but insufficient because it cannot by itself evidence configured use. This is why RAIDT defines the run as the unit of governance and introduces a run-level evidence pack as the proof object for material use.

RAIDT therefore extends Responsible AI rather than rejecting it. Its five pillars (Responsibility, Auditability, Interpretability, Dependability, Traceability) translate broad normative commitments into an inspectable score profile for one run. The anchors 1=missing / 3=partial / 5=audit-ready make the quality of evidence explicit. Crucially, RAIDT also treats influence methods as governance interventions, because retrieval, alignment, prompting, or adaptation can alter both behaviour and what evidence exists for later review.

Consider a public-service eligibility adviser using GenAI to interpret a benefits rule for one claimant. A conventional Responsible AI approach may ensure that the organisation has an ethics policy, model documentation, and a general review routine. If the claimant later challenges the advice, those artefacts still may not show which prompt was used, which policy clause was retrieved, what version of the rule base was active, or whether a human checker approved the answer.

Under RAIDT, that single interaction is governed as a run. The run-level evidence pack records the prompt, model version, retrieval snapshot, document identifiers, hashes, output, and review action. The resulting score profile can then show whether the case was genuinely auditable and traceable. The difference is not rhetorical. It is the difference between claiming responsible practice in the abstract and being able to reconstruct one contested use in evidence.

• 11-RAIDT_Academic_Logic_M_v11
• 15-RAIDT-IS-Governance_M_v07