• qa_deck_100#slide 50 · Outputs, review decisions, and retention

A review decision adds governance evidence that the generated output alone cannot supply. Across the papers, RAIDT is explicit that governance cannot rest on the text produced by a model, because a single output does not show which controls were active, who reviewed it, whether it was approved under the right conditions, or how responsibility was exercised in context. The run-level evidence pack therefore records not only inputs, configuration provenance and outputs, but also oversight and decision recording. In practice, that means preserving what humans reviewed, what they changed, what they approved, and what was escalated, so that the organisation can later reconstruct whether the output was actually relied upon and under what judgement.

This matters because RAIDT treats the run as the unit of governance, not the model or the document in isolation. A review decision converts a generated artefact into a reviewable organisational event. It shows whether the output remained a draft, became an accepted record, was edited before use, was rejected, or triggered escalation or incident handling. In RAIDT terms, that additional layer supports the five pillars (Responsibility, Auditability, Interpretability, Dependability, Traceability) by linking system behaviour to human control. It also enables a meaningful score profile, because the scored object is the run-level evidence pack rather than the output text alone.

Taken together, the papers imply that a review decision is the difference between content and accountable use. Without it, an organisation may possess a plausible answer yet still be unable to show contestability, allocate responsibility, or justify higher anchors 1=missing / 3=partial / 5=audit-ready for governance readiness.

In a hospital discharge-summary workflow, a GenAI assistant drafts a note from clinician inputs. The draft itself may read fluently, but that does not establish whether it was safe to enter into the patient record. A review decision adds the missing governance layer: the clinician reviewed the draft, corrected a medication instruction, approved the final text, and logged why the change was necessary.

If the draft had instead omitted a red-flag symptom, the reviewer might reject it and link the run to an incident ticket for further investigation. In RAIDT terms, the same output family would then lead to very different governance conclusions because the run-level evidence pack would show different oversight actions. The decision record therefore turns a generated note into a reviewable clinical event rather than leaving it as an uncontextualised piece of text.

• 08-RAIDT_Foundations_M_V50
• 13-RAIDT-Evidence-Review_M_v10
• 18-RAIDT-Technical-Foundation_M_v04