• integrated_82#Q2.7

Principles, policies, model cards, system cards, and periodic lifecycle controls are not enough on their own because they usually describe intended properties, governance arrangements, or system-level characteristics, whereas GenAI risk often emerges in the specifics of a particular run. The papers repeatedly stress that a model may be well documented and an organisation may have strong policies, yet a contested output still cannot be reconstructed if the prompt, retrieval context, configuration, tool outputs, oversight steps, and version identifiers from that use were never preserved. In that sense, these familiar artefacts are necessary but insufficient: they provide context, but not case-level evidence.

They are also often static or episodic in relation to a system that is dynamic at run time. Periodic audits and lifecycle controls may confirm that a governance process exists, but they can miss how small changes in prompt structure, retrieval behaviour, alignment settings, or adapters alter outputs in practice. The papers warn that governance expressed only as policy or checklist can drift into compliance theatre. RAIDT therefore complements these artefacts by treating the run as the unit of governance, requiring a run-level evidence pack, and assessing the resulting evidence through a score profile. This is why model cards remain useful but cannot, by themselves, satisfy the evidentiary demands of Responsibility, Auditability, Interpretability, Dependability, and Traceability for a concrete GenAI use.

A bank uses GenAI to draft an adverse-action explanation after a credit refusal. The bank may possess a model card, internal fairness policy, and quarterly control review, yet a customer challenge turns on a narrower question: what exactly informed this explanation? If the run record does not preserve the prompt template, reason-code mapping, versioned decision criteria, and reviewer edits, the institution cannot show whether the explanation was grounded in the actual decision record or embellished by the model.

RAIDT changes that by requiring the explanation to be tied to a run-level evidence pack and then scored, particularly for Interpretability and Traceability. The existing governance artefacts still matter, but they become supporting context rather than substitutes for case-specific evidence.

• 08-RAIDT_Foundations_M_V50
• 11-RAIDT_Academic_Logic_M_v11