• qa_deck_100#slide 14 · Boundaries, scope, and limitations

RAIDT's boundary conditions specify the situations in which the framework is theoretically strongest and practically most useful. Across the papers, the core boundary is organisational GenAI use where outputs materially influence decisions, records, communications, or services and where later reconstruction, review, or challenge is plausible. In that setting, RAIDT treats the run as the unit of governance, because risk materialises in one configured use in context rather than in the model abstractly. The central artefact is therefore the run-level evidence pack, which makes one material use inspectable through the five pillars (Responsibility, Auditability, Interpretability, Dependability, Traceability). The resulting score profile is intended to support governance readiness assessment through anchors 1=missing / 3=partial / 5=audit-ready.

The framework is especially well suited to repeated organisational use, use across teams, and high-impact workflows in which actions must be justified. It also assumes that organisations can capture and preserve bounded run evidence under suitable access, retention, and privacy controls. This includes prompts, configurations, retrieved context where relevant, outputs, checks, and oversight records. By contrast, RAIDT is less relevant for trivial or ephemeral uses and cannot compensate where evidence cannot lawfully or technically be captured. The papers also emphasise proportionality: not every use needs the same depth of instrumentation, and RAIDT does not guarantee correctness or replace legal, ethical, or professional judgement. Its boundary conditions therefore discipline its claims: RAIDT is a bounded governance design for reviewable, consequential organisational GenAI use, not a universal theory of all AI governance.

A public-service agency uses GenAI to draft eligibility advice for housing support. This is within RAIDT's boundary conditions because the output influences a service decision, may be challenged later, and must be reviewable by supervisors, auditors, and case workers. A suitable run-level evidence pack would record the prompt template, model deployment, retrieval snapshot of the exact policy clauses used, the generated advice, and the human review outcome.

In that scenario, RAIDT can score the run across the five pillars (Responsibility, Auditability, Interpretability, Dependability, Traceability) and surface whether the advice is governable rather than merely fluent. If the agency cannot preserve the policy snapshot because of a broken logging pipeline, the case may still be scored, but likely with low auditability or traceability. That illustrates why the framework's boundary conditions matter: RAIDT is strongest where evidence can be preserved and where review is institutionally significant.

• 08-RAIDT_Foundations_M_V50
• 12-RAIDT_DSR_Theory_M_v8