• qa_deck_100#slide 31 · Design science and mid-range design theory

RAIDT brings design science, mid-range theory, and mechanism-based explanation into a single cumulative research logic. Design science provides the programme structure: RAIDT starts from a concrete organisational problem, namely the inability to reconstruct and contest a specific GenAI use event, and responds by designing an artefact set. That artefact set is the run-level evidence pack plus a score profile built around the five pillars (Responsibility, Auditability, Interpretability, Dependability, Traceability), scored through anchors 1=missing / 3=partial / 5=audit-ready. In this sense, RAIDT is not a principles-only proposal; it is a design response intended for repeated evaluation, calibration, and refinement across domains.

Mid-range theory gives RAIDT the right level of abstraction. The theory is broader than a local workflow rule, yet narrower than a universal theory of AI governance. It travels across healthcare, finance, public service, cybersecurity and HR because it preserves the same core constructs while stating explicit boundary conditions. Mechanism-based explanation then clarifies why outcomes change. RAIDT treats influence methods as governance interventions: structured prompting, retrieval augmentation, PEFT/LoRA, and preference-based alignment affect both system behaviour and the evidentiary record. When these mechanisms are instrumented and logged, governance artefacts make run as the unit of governance inspectable; when they are opaque or weakly captured, auditability, traceability and other dimensions weaken. The design science programme therefore builds and evaluates the artefact, while the mid-range, mechanism-based theory explains how and under what conditions it produces observable governance outcomes.

In a public-service eligibility advice workflow, a council uses GenAI to draft a benefits explanation for a claimant. A design science approach would compare several configurations: baseline prompting, retrieval-augmented generation over current policy, and a stacked setup with retrieval plus structured review prompts. For each run, the team captures a run-level evidence pack and produces a score profile across the five pillars (Responsibility, Auditability, Interpretability, Dependability, Traceability), using anchors 1=missing / 3=partial / 5=audit-ready.

The mid-range theory then explains the results rather than merely describing them. If the retrieval-based configuration improves Traceability and Auditability, the mechanism is not simply that the AI 'performed better'; it is that retrieval snapshots, document hashes, and constrained prompts made the case reconstructable. If an alignment layer improves Responsibility but reduces Interpretability, RAIDT reads that as a governance trade-off caused by a specific intervention. This is how repeated runs become cumulative governance knowledge rather than isolated technical tests.

• 11-RAIDT_Academic_Logic_M_v11
• 12-RAIDT_DSR_Theory_M_v8