• workshop_dense_100#slide 34

A mid-range design theory sits between grand theory and isolated description. It explains a bounded class of socio-technical problems by specifying the constructs, mechanisms, propositions, and boundary conditions through which an artefact is expected to work. In RAIDT, the bounded problem is not AI governance in general but governance of organisational GenAI runs that may later need reconstruction and challenge. The theory therefore keeps its scope disciplined: RAIDT defines the run as the unit of governance, centres the run-level evidence pack as the proof object, and renders governance readiness through a score profile across the five pillars (Responsibility, Auditability, Interpretability, Dependability, Traceability), with anchors 1=missing / 3=partial / 5=audit-ready.

RAIDT itself is the clearest example. It proposes that when organisations preserve run-level evidence and treat influence methods as governance interventions rather than merely technical tweaks, they can produce more reviewable and comparable outcomes. Structured prompting, retrieval augmentation, or PEFT/LoRA may strengthen some pillars, but only if the evidence required for reconstruction is also preserved. This matters because RAIDT becomes cumulative and testable. Researchers can compare runs, sectors, and configurations; practitioners can redesign controls when one pillar persistently underperforms. Without a mid-range design theory, RAIDT would risk becoming either an over-ambitious universal claim or a static checklist with no explanatory power.

Consider an HR shortlist-justification workflow in which GenAI helps explain why candidates were or were not advanced. RAIDT treats that one shortlist explanation as the governed object, not the recruitment system in the abstract. The run-level evidence pack would preserve the criteria, prompt template, model deployment, adapter lineage where PEFT/LoRA is used, output, and reviewer sign-off. The score profile then reveals whether the organisation can actually justify and reconstruct the decision.

This matters because influence methods as governance interventions may alter the result. A domain-tuned adapter might improve Dependability for HR language, yet Traceability could still fail if version identifiers or evidence links are missing. Mid-range design theory helps the organisation learn from that pattern instead of merely ticking a recruitment checklist.

• 12-RAIDT_DSR_Theory_M_v8
• 11-RAIDT_Academic_Logic_M_v11
• 17-RAIDT-Sociotechnical_M_v6