• qa_deck_100#slide 9 · Evidence pack and score profile

The run-level evidence pack is the standard proof object that RAIDT uses to govern GenAI in organisational work. It is a structured record of what happened in one configured use, at one time, for one task, in one context. This matters because RAIDT shifts attention away from abstract claims about a model and towards reconstructable evidence of an actual use event. In that sense, the run-level evidence pack operationalises the run as the unit of governance: a run is no longer just an interaction, but an inspectable artefact that can be reviewed by an auditor, manager, domain supervisor, compliance officer, or internal quality reviewer who was not present when the output was produced.

Conceptually, the pack links evidence capture to measurement. It is the scored object from which RAIDT derives a score profile across the five pillars (Responsibility, Auditability, Interpretability, Dependability, Traceability). The papers therefore present the pack not as administrative paperwork, but as the evidentiary hinge between principle and judgement. It records prompts, configurations, outputs, retrieval context where used, checks, and oversight decisions, while also making space for retention metadata, access control, and organisational linkage to risk registers or incident logs. This is why RAIDT also treats influence methods as governance interventions: prompting, retrieval augmentation, PEFT or LoRA, and alignment controls must be logged because they change both system behaviour and what counts as adequate evidence under the anchors 1=missing / 3=partial / 5=audit-ready.

In the healthcare vignette, a clinician asks GenAI to summarise a chest-pain consultation into Symptoms, Diagnosis, Treatment, and Red Flags. The run-level evidence pack preserves the constrained system role, the instruction that unsafe invention is banned, any uncertainty statement requested, the model and prompt versions, and the clinician oversight rule. If the clinic later investigates whether a warning sign was omitted, reviewers do not have to rely on the final summary alone.

They can inspect the run-level evidence pack, see what constraints were active, and decide whether the run deserved a stronger Responsibility or Interpretability score. The pack therefore functions both as a reconstruction record and as the basis for disciplined review.

• 08-RAIDT_Foundations_M_V50
• 11-RAIDT_Academic_Logic_M_v11