• qa_deck_100#slide 7 · What RAIDT is and what it governs

Run-level governance in RAIDT means governing each material use of generative AI through a bounded evidentiary object rather than through policy statements alone. In practice, this requires instrumentation and workflow design so that every governed run produces a run-level evidence pack. The papers describe this pack as recording context-of-use, configuration provenance, inputs and outputs, integrity markers such as hashes, oversight actions, and any post-run monitoring signals relevant to review. The purpose is to convert scattered telemetry and fragmented logs into a reviewable governance record for one configured use.

Operationally, this also means treating influence methods as governance interventions. Prompt structure, retrieval augmentation, PEFT or LoRA adapters, alignment layers, and review checkpoints are not merely engineering choices; they are governed configuration choices whose effects on governance outcomes must be visible. Reviewers then assess the run through the five pillars (Responsibility, Auditability, Interpretability, Dependability, Traceability) and produce a score profile with anchors 1=missing / 3=partial / 5=audit-ready. That profile can support deployment gating, escalation to human review, repeat-run testing for dependability, and targeted improvement where evidence is weak. Programme-level audits still matter, but in RAIDT they become stronger because they can sample and compare actual runs rather than inferring quality from policies or model descriptions alone.

In a public-service eligibility workflow, run-level governance would mean storing the exact policy clause or retrieved rule text used for the advice, its version, the prompt template identifier, the model and settings, the generated explanation, and the caseworker's decision to accept, amend, or escalate. If the explanation is later disputed, reviewers do not rely on a generic compliance statement.

Instead, they inspect that run-level evidence pack and score whether the advice was auditable, interpretable, and traceable. If repeated runs show weak auditability, the organisation can tighten retrieval capture or require stronger human review before advice is issued. That is run-level governance as an operating practice, not merely as an abstract principle.

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