• qa_deck_100#slide 20 · Why current governance leaves a run-level gap
• integrated_82#Q2.6

RAIDT addresses a residual literature gap between broad responsible AI commitments and the evidence needed to govern a single organisational GenAI use. Across responsible AI principles, model cards, datasheets, audit routines, explainability methods, and MLOps observability, the literature specifies many desirable properties and many logging practices, but it does not converge on a standard proof object for one configured run in context. The papers argue that this matters because generative AI risk materialises at run time: prompts, retrieval sources, model deployments, adapters, decoding settings, tool calls, and oversight decisions can all change what happened in a particular case.

RAIDT responds by defining the run as the unit of governance and by specifying a run-level evidence pack that can be reconstructed, reviewed, compared, and challenged. The pack records the artefacts needed to inspect one use event, while the score profile interprets those artefacts across the five pillars (Responsibility, Auditability, Interpretability, Dependability, Traceability). In this design, influence methods as governance interventions must themselves be evidenced, because structured prompting, retrieval augmentation, PEFT/LoRA, or alignment controls reshape both behaviour and accountability. The literature gap is therefore not simply a lack of logs; it is the absence of a bounded, governance-grade evidence object that turns claims such as responsible use into inspectable evidence, with anchors 1=missing / 3=partial / 5=audit-ready.

A local authority uses retrieval-augmented GenAI to draft eligibility advice for housing support. The service has a policy pack, a model card, and general audit procedures, and the output even shows source citations. A complaint arrives months later from a citizen who says the advice relied on an outdated policy clause. If the authority has no retrieval snapshot, no prompt version, and no record of who approved the message, it cannot reconstruct the configured run even though it can show general governance documents.

RAIDT closes that gap by requiring a run-level evidence pack for that case: the exact policy clause version, retrieval snapshot identifiers and hashes, prompt template ID, model deployment, output hash, and reviewer action. The disputed case can then be scored through a score profile, making weak auditability or traceability visible before similar failures recur.

• 08-RAIDT_Foundations_M_V50
• 13-RAIDT-Evidence-Review_M_v10
• 16-RAIDT-Audit-Accountability_M_v05