• qa_deck_100#slide 26 · Uncertainty, hallucination, and over-reliance

RAIDT closes the compliance-measurement gap by supplying what the papers call an operational bridge, or measurement hinge, between high-level obligations and inspectable evidence. Legal instruments and standards often say that organisations should maintain records, support oversight, provide intelligible explanations, monitor robustness, and preserve provenance. What they often do not provide is a common measurement object that shows whether those expectations were met in a particular case. RAIDT answers that problem with the run-level evidence pack and a score profile for the five pillars (Responsibility, Auditability, Interpretability, Dependability, Traceability). In practice, the rubric translates governance requirements into evidence fields and anchored judgements, using anchors 1=missing / 3=partial / 5=audit-ready to make readiness reviewable rather than rhetorical.

The framework does not over-claim because the papers explicitly delimit what is being measured. RAIDT measures governance readiness from preserved evidence; it does not certify factual correctness, guarantee safety, or prove legal compliance by itself. It does not replace clinical judgement, formal public-sector decision procedures, or finance-specific legal duties. Instead, it makes the evidentiary basis of governance visible, calibrates scoring through predefined anchors and reviewer judgement, and supports standards mapping without pretending that a score abolishes substantive domain responsibility. That restraint is central to the framework's credibility: RAIDT is strong precisely because it claims an auditable measurement method, not a universal proof of trustworthy AI.

In a healthcare note-summarisation pilot, a trust may need to show that human oversight was active, unsafe invention was constrained, uncertainty was communicated, and the output could be reviewed later. A policy manual can state those expectations, but it does not measure whether they were met in one consultation-summary run. RAIDT closes that gap by capturing the prompt, safety constraints, model version, output, and clinician review within a run-level evidence pack, then producing a score profile.

If the run scores strongly on Responsibility and Auditability, the organisation has concrete evidence for internal assurance. If it scores poorly, the same method reveals where governance is thin. Crucially, RAIDT still does not claim the summary is medically correct simply because the paperwork is complete; it claims only that the governance of that run is inspectable and measurable.

• 00-RAIDT_Wording_v2
• 11-RAIDT_Academic_Logic_M_v11