• integrated_82#Q1.2

In RAIDT, 'run-level' means that governance evidence is attached to an individual use event rather than to the GenAI system in the abstract. The focus is not merely whether an organisation has a policy, a model card, or a general assurance claim. Instead, the question is whether a particular configured use can be reconstructed, examined, and justified after the fact. Run-level governance therefore centres on what happened in that specific run: which prompt template was used, which model deployment and parameters were active, what sources were retrieved, what output was produced, and what checks or oversight followed.

This is why the run-level evidence pack is the central object in RAIDT. It turns a moment of system use into a reviewable proof object. The evidence from that one run is then scored across the five pillars (Responsibility, Auditability, Interpretability, Dependability, Traceability), producing a score profile for the run rather than a generic judgement about the whole system. The scoring logic is expressed through anchors 1=missing / 3=partial / 5=audit-ready, so run-level means that evidence quality is judged where GenAI risk actually appears.

Run-level also means that RAIDT studies influence methods as governance interventions. Structured prompting, retrieval augmentation, PEFT or LoRA, and preference-based alignment are not treated as informal engineering choices. They are governed configuration choices that must be logged because they affect both behaviour and what evidence can be captured. In practice, then, 'run-level' names the operational granularity for comparison, audit sampling, repeat-run testing, and improvement.

In the cybersecurity vignette, a security analyst asks GenAI to triage an alert and recommend next steps. A run-level view does not stop at the recommendation text. It records the prompt structure, the approved threat-intelligence source used for retrieval, the retrieved snapshot identifiers, the model configuration, and the repeat-run tests used to check variance and drift.

This makes the recommendation governable at the level that matters operationally. If two similar alerts produce materially different advice, reviewers can inspect whether the difference came from prompt variation, retrieval differences, or configuration drift. Without a run-level record, the organisation sees only fluent outputs; with run-level evidence, it can score dependability and auditability on what actually occurred.

• 08-RAIDT_Foundations_M_V50
• 11-RAIDT_Academic_Logic_M_v11