• qa_deck_100#slide 16 · Boundaries, scope, and limitations

The main limitations of RAIDT follow directly from its design scope. RAIDT is a bounded governance framework for organisational generative AI use, not a truth-guaranteeing or risk-eliminating system. Its purpose is to make one configured use inspectable through a run-level evidence pack and a score profile, with the run as the unit of governance. Accordingly, RAIDT can show whether evidence exists for the five pillars (Responsibility, Auditability, Interpretability, Dependability, Traceability), but it cannot itself prove that an output is factually correct, legally sufficient, or professionally safe. Correctness remains task-specific and may still require domain expertise, external verification, or additional controls. The papers are explicit that RAIDT does not replace law, regulation, domain safety practice, or expert judgement.

A second limitation is conditionality. RAIDT is strongest where organisations can preserve bounded run evidence under appropriate access, retention, and privacy controls. If prompts, retrieved context, tool traces, reviewer decisions, or configuration identifiers are not captured, the framework can reveal governance weakness but cannot compensate for missing evidence by design alone. The same applies to trade-offs: more instrumentation may strengthen auditability while increasing burden; stronger alignment may support responsibility while complicating interpretability. This is why the score profile and its anchors 1=missing / 3=partial / 5=audit-ready should be read as indicators of governance readiness, not as a blanket claim that the underlying system is correct, harmless, or universally suitable across all contexts.

In healthcare note summarisation, a hospital may use a GenAI assistant to draft a chest-pain discharge summary. RAIDT can require a run-level evidence pack containing the prompt version, retrieval snapshot from the internal guideline corpus, output hash, and clinician review decision. A strong score profile may show that the run was well governed: the evidence is reconstructable, uncertainty was stated, and oversight was recorded.

Even so, RAIDT cannot certify that the clinical content is medically correct. If the retrieved guideline is outdated, if the symptoms indicate an unusual presentation, or if the clinician overlooks a red flag, the framework has not removed domain risk. What it has done is make the limitations visible and reviewable. That matters because the organisation can challenge the run, inspect what happened, and improve controls without pretending that governance evidence is the same thing as clinical truth or legal certification.

• 08-RAIDT_Foundations_M_V50
• 11-RAIDT_Academic_Logic_M_v11
• 12-RAIDT_DSR_Theory_M_v8