• integrated_82#Q3.16

RAIDT retains a five-pillar profile rather than a single score because governance is treated as a multi-objective design problem. The five pillars (Responsibility, Auditability, Interpretability, Dependability, Traceability) together define governance readiness, but the framework does not assume they move together. The foundations paper explicitly notes that improving one dimension can strengthen or weaken another. Retrieval grounding may improve traceability while introducing privacy or intellectual-property risk; prompt constraints may improve interpretability while reducing helpfulness; alignment layers may strengthen responsibility while reducing transparency if their policies are not documented. A single score would suppress these interactions instead of exposing them.

The profile also protects construct clarity. RAIDT is designed to operationalise governance claims as inspectable evidence rather than narrative assurance. When reviewers score a run-level evidence pack against anchors 1=missing / 3=partial / 5=audit-ready, they are not certifying overall goodness; they are judging whether evidence exists for specific governance questions. If the five pillars were collapsed into one number too early, the measure would lose the ability to distinguish where the run failed and why.

Finally, the five-pillar profile supports organisational learning across sectors. Healthcare, finance, public services, cybersecurity, HR, and education may set different thresholds, yet they can still inspect the same score profile and compare runs on a common basis. RAIDT therefore permits a composite mean, but keeps the five-pillar profile as the main result because governance decisions depend on the pattern, not only the average.

In cybersecurity alert triage, a GenAI assistant may log every prompt, tool call, and retrieval snapshot, so Auditability and Traceability are strong. It may also route severe cases for analyst sign-off, supporting Responsibility. Yet if small prompt changes cause materially different prioritisation advice across repeated runs, Dependability is weak.

A single score of 3.8 could be read as good enough. The five-pillar profile shows a different picture: the workflow is well logged and reviewable, but not yet stable enough for routine operational reliance. That matters because the right intervention is not more documentation; it is repeat-run testing, configuration control, and tighter perturbation tolerances. The profile therefore turns governance into targeted improvement rather than headline reporting.

• 08-RAIDT_Foundations_M_V50
• 00-RAIDT_Scoring_v1
• 13-RAIDT-Evidence-Review_M_v10