• qa_deck_100#slide 2 · RAIDT core and mindmap orientation

RAIDT makes the run the key governance unit because organisational accountability questions usually arise around one configured use in context, not around a model in the abstract. Across the papers, the central argument is that generative AI risk materialises at run time: prompts, prompt templates, retrieval sources, tool calls, decoding settings, adaptation layers, alignment controls, and human oversight all shape the output that enters work. The same model family can therefore produce materially different outputs under different configurations. In that setting, model-level documentation remains useful, but it cannot show what actually happened when a disputed summary, recommendation, or explanation was produced.

For that reason, RAIDT adopts run as the unit of governance and treats governance as an evidence problem. A run-level evidence pack records the purpose and stakes of the task, configuration provenance, input and output records, integrity markers, and oversight decisions for one configured use. That bounded record is then assessed through the five pillars (Responsibility, Auditability, Interpretability, Dependability, Traceability) and expressed as a score profile using anchors 1=missing / 3=partial / 5=audit-ready. This makes governance readiness inspectable and comparable across runs, supports reconstructability and contestability, and allows audits or reviews to focus on real uses rather than on narrative assurance alone. In RAIDT terms, the run matters because it is the level at which a specific output can later be reconstructed, reviewed, challenged, and improved.

Consider the healthcare discharge-summary scenario used in the foundations paper. A hospital compares two configurations of the same underlying model: one uses a free-form prompt, while another uses a structured discharge template, an uncertainty statement, retrieval from internal clinical guidance, and clinician review before filing the note. If a patient later challenges an omission, the relevant governance question is not whether the model was generally documented. It is whether this run can be reconstructed.

A run-level evidence pack for that specific case would preserve the prompt hash, retrieval snapshot identifiers, model version, output hash, and the clinician's approval or amendment. Reviewers can then score the run across the five pillars and see whether the configuration was governed adequately. That is why RAIDT treats the run, rather than the model alone, as the object of governance.

• 08-RAIDT_Foundations_M_V50
• 16-RAIDT-Audit-Accountability_M_v05