• integrated_82#Q3.2

In RAIDT, the run as the unit of governance means the run-level evidence pack must contain the minimum metadata needed to reconstruct one configured use in context, rather than a generic description of the model or workflow. At minimum, that includes a stable run ID and timestamp; the context of use such as purpose, domain, decision type, user role, intended audience, and applicable constraints; the full prompt or template with version identifiers and system instructions; the model provider, deployment or version identifier, and decoding or configuration settings; and any enabled tools or retrieval settings. Because influence methods as governance interventions materially shape behaviour, retrieval queries, snapshot identifiers, document hashes, tool outputs, and source pointers belong in the pack where relevant.

The pack must also preserve the governed result and the review trail. That means the produced output, an output hash or immutable storage pointer, and the human or automated checks applied to the run, including reviewer role, approvals, edits, exceptions, escalations, and incident flags where these occurred. Where privacy or intellectual-property constraints prevent full content storage, the minimum remains immutable identifiers, hashes, time-stamped snapshots, retention metadata, and access controls that still permit later verification. This is why the run-level evidence pack is the scored object for the five pillars (Responsibility, Auditability, Interpretability, Dependability, Traceability) and its score profile: under anchors 1=missing / 3=partial / 5=audit-ready, absent metadata directly weakens Auditability and Traceability because the run cannot be reconstructed or independently reviewed.

In a healthcare note summarisation workflow, a hospital should preserve a run-level evidence pack for each materially used draft. A minimally adequate pack would record the run ID and timestamp, the discharge-summary prompt template version, the model deployment ID, decoding settings, any retrieved guideline snapshot ID and hash, the generated summary, an output hash, and the clinician safety-review flag. If the patient note itself cannot be widely stored, the organisation can still retain controlled pointers, immutable identifiers, and access-controlled hashes.

That package allows a later reviewer to see exactly which governed run produced the summary that entered the patient record, what evidence sources were active, and what review happened before sign-off. Without those minimum metadata fields, the hospital would have only a general policy or model card, not a reconstructable run-level evidence pack.

• 08-RAIDT_Foundations_M_V50
• 13-RAIDT-Evidence-Review_M_v10