• workshop_table17#tag-band S3 · 40–65 min

RAIDT treats the run as the governance unit because the practical governance problem is not simply whether a model exists within a controlled programme, but whether one configured use can later be shown, assessed, and justified. The papers repeatedly argue that generative AI behaviour is shaped by runtime configuration: prompts, templates, retrieved context, tools, model versions, decoding choices, alignment layers, and oversight actions jointly produce the output that becomes consequential. As a result, governing the model alone leaves an evidentiary gap whenever a later review concerns a particular note, explanation, shortlist, or advisory message.

Run-level evidence solves that gap by turning a configured use into a bounded proof object. The run-level evidence pack records intent, configuration provenance, inputs and outputs, integrity markers, and human decisions for one run, while the score profile evaluates the five pillars (Responsibility, Auditability, Interpretability, Dependability, Traceability) through anchors 1=missing / 3=partial / 5=audit-ready. This solves several governance problems at once: reconstructability after incidents, comparability across configurations, contestability for affected parties, and organisational learning from repeated weak scores. It also separates audit-ready evidence from raw logging: RAIDT is not interested in data exhaust alone, but in an evidence object that broader audits, assurance exercises, and standards alignment can sample and interrogate. In short, run-level evidence solves the missing operational link between governance intention and reviewable organisational proof.

In a healthcare note-summarisation workflow, a hospital may already have a policy, a model card, and a periodic audit programme. Yet if a discharge note is later alleged to have omitted a critical risk, those artefacts do not by themselves show what happened in that use. The organisation needs the run-level evidence pack for the disputed note: prompt version, retrieval snapshot, model and parameter settings, output hash, and clinician review action.

That evidence allows the hospital to reconstruct the case, score the run across the five pillars, and identify whether the weakness lay in prompting, retrieval, oversight, or broader workflow design. This is the specific governance problem that run-level evidence is designed to solve.

• 08-RAIDT_Foundations_M_V50
• 13-RAIDT-Evidence-Review_M_v10
• 16-RAIDT-Audit-Accountability_M_v05