• qa_deck_100#slide 35 · Socio-technical governance and accountability lineage

In RAIDT, the evidence object is the run-level evidence pack derived from a run record. It is a bounded, structured, versioned, and reviewable bundle of records for one configured use event. The papers describe it as the minimum evidentiary bundle needed to reconstruct and review what happened: run ID and timestamp, task context, prompt or template and version, model deployment and settings, tool or retrieval configuration, retrieved snapshot identifiers and hashes where relevant, the output and integrity hash, and the human or automated checks applied.

Its governance role is specific. RAIDT treats the run as the unit of governance, so the scored object is not a general policy statement, a model card, or a vendor factsheet. It is the run-level evidence pack, which is then interpreted through a score profile across the five pillars (Responsibility, Auditability, Interpretability, Dependability, Traceability). This is why the evidence object is more than logging: it packages records so that a reviewer can inspect, compare, and contest one run in context. It also captures influence methods as governance interventions, since prompting, retrieval augmentation, PEFT/LoRA, and preference-based alignment change both system behaviour and what evidence must exist. In practical use, RAIDT makes governance readiness visible through anchors 1=missing / 3=partial / 5=audit-ready.

In a healthcare note summarisation workflow, a GenAI tool drafts a discharge summary for a clinician. Under RAIDT, the evidence object is not just the final summary. It includes the prompt template ID, the model deployment identifier, decoding settings, any retrieval snapshot hash, the output text and output hash, plus the safety check and oversight flag showing whether a clinician reviewed or escalated the draft.

That package matters because the hospital may later need to ask whether uncertainty was communicated properly, whether the right retrieval context was used, and whether the summary was reviewed before entering the record. With a complete run-level evidence pack, an independent reviewer can reconstruct the run and score it rather than rely on memory or informal explanation.

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