• qa_deck_100#slide 88 · Gating, monitoring, review, and corrective action

RAIDT frames corrective action after weak runs or failed reviews as an evidence-led governance response rather than as ad hoc troubleshooting. Because RAIDT treats the run as the unit of governance, the immediate object of review is the run-level evidence pack and its score profile across the five pillars (Responsibility, Auditability, Interpretability, Dependability, Traceability). In this framing, a poor outcome is not reduced to whether the model answered badly; it is interpreted through evidence quality, oversight quality, and reconstructability. The rubric's anchors 1=missing / 3=partial / 5=audit-ready make weak runs legible: low scores indicate missing evidence, incomplete controls, or unstable behaviour that cannot yet support justified organisational reliance.

The papers therefore position corrective action as a structured move from review to monitoring and control updates. RAIDT explicitly links scoring to deployment gating, escalation, and remediation. Low Auditability or Traceability scores call for instrumentation fixes such as stronger logging, prompt version control, preserved retrieval snapshots, hashes, and better evidence retention. Low Responsibility scores call for stronger constraints, clearer escalation guidance, or added human review. Low Dependability scores call for repeat-run testing, configuration stabilisation, and closer monitoring. This logic matters because RAIDT treats influence methods as governance interventions: prompts, retrieval, adapters, and alignment layers are governed artefacts that can be changed after weak evidence or incidents. Corrective action is therefore how RAIDT turns review findings into organisational learning, post-incident reconstruction, and more governable future runs rather than leaving governance at the level of policy statements alone.

In a cybersecurity alert triage workflow, a GenAI assistant recommends downgrading an alert after reviewing retrieved logs and threat notes. The run-level evidence pack shows that the alert was produced under a prompt template that did not force uncertainty reporting, and the retrieval context was not stored with stable identifiers. Reviewers therefore record a weak score profile: Dependability is low because repeat runs vary, and Traceability is low because the evidence cannot show exactly which passages informed the recommendation.

RAIDT would frame corrective action in operational terms. The team would revise the prompt to require explicit uncertainty statements, preserve retrieval snapshots and hashes, and introduce a human review checkpoint for low-confidence triage runs. The next review would test whether those changes raise the relevant pillar scores and produce a more audit-ready run record.

• 08-RAIDT_Foundations_M_V50
• 18-RAIDT-Technical-Foundation_M_v04