• workshop_dense_100#slide 35

In RAIDT, mechanism-based explanation means explaining governance outcomes by identifying the recurrent processes through which artefacts and interventions produce reviewable organisational use. It is therefore more precise than saying that governance should be responsible or transparent. RAIDT defines the run as the unit of governance, uses the run-level evidence pack as the bounded proof object for one configured use, and expresses governance readiness through a score profile across the five pillars (Responsibility, Auditability, Interpretability, Dependability, Traceability). Mechanism-based explanation asks how those outcomes are generated. The answer is that prompts, retrieval layers, adapters, alignment controls, logging routines, review checkpoints, and evidence-retention practices jointly shape what can later be reconstructed, challenged, and relied upon.

This matters in RAIDT for three reasons. First, it gives the framework a defensible academic logic: the papers present RAIDT not as a checklist, but as a mechanism-based mid-range design theory whose claims can be examined and refined. Second, it makes scoring operational. Because the scored object is the run-level evidence pack, the score profile can use anchors 1=missing / 3=partial / 5=audit-ready without collapsing into opinion or narrative assurance. Third, it matters organisationally because shared evidence objects change governance routines. Managers can judge reliance more carefully, auditors can sample runs, compliance teams can inspect preserved evidence, and affected people gain a more credible basis for challenge. In short, mechanism-based explanation matters because it links design choices to governance consequences at the point where organisational risk actually materialises.

A hospital uses GenAI to draft discharge summaries from clinician notes. This is a useful RAIDT example because the output enters the patient record and may later be contested. If the hospital keeps only the final text, governance remains largely narrative: staff may say that review occurred, but they cannot show exactly what prompt was used, whether retrieval supported the draft, or how uncertainty was handled.

Under RAIDT, the run-level evidence pack stores the prompt template ID, model deployment ID, any retrieval snapshot, the output, a safety check, and the clinician oversight flag. Structured prompting and explicit escalation language act as mechanisms for Responsibility and Interpretability; logging, hashes, and reviewer records act as mechanisms for Auditability and Traceability. The score profile then shows whether this particular run is weak, partial, or audit-ready. That matters because if a discharge summary is challenged later, the hospital can reconstruct one use event in context rather than relying on memory, screenshots, or broad claims that the system was governed.

• 11-RAIDT_Academic_Logic_M_v11
• 17-RAIDT-Sociotechnical_M_v6