• qa_deck_100#slide 30 · Design science and mid-range design theory

Mechanism-based explanation adds something more demanding than description to RAIDT. It does not merely name desirable governance values or list the five pillars (Responsibility, Auditability, Interpretability, Dependability, Traceability). Instead, it explains how a run-level evidence pack, a score profile, and influence methods as governance interventions are expected to generate observable governance outcomes when the run as the unit of governance is instrumented properly. In the theory paper, this is the difference between a useful framework and a design theory: RAIDT becomes able to say why governance readiness rises or falls, under what conditions, and through which recurrent socio-technical processes. That explanatory layer is important because configured generative AI use is shaped by prompts, retrieval, tooling, review steps, and retention practices, not by model capability in isolation.

What mechanism-based explanation therefore adds is causal and cumulative logic. It links structured prompting to improved interpretability, preserved retrieval snapshots to stronger traceability, logging and hashing to auditability, and versioned adaptation or alignment controls to dependability and responsibility. It also explains why these gains are conditional: retrieval without preserved snapshots may look grounded yet remain weakly auditable; extensive instrumentation may improve reviewability while increasing burden. In RAIDT, the mechanism-based move turns governance from narrative assurance into inspectable design knowledge. That makes the framework stronger academically, because its claims become testable, and stronger practically, because managers, auditors, and reviewers can examine how particular artefacts and interventions changed governance outcomes in one run rather than only asserting that governance existed in general.

Consider a public-service eligibility workflow in which a staff member uses GenAI to interpret benefit rules for one claimant. Without RAIDT, the organisation may keep the advice text but not the exact rule version, retrieved clause, prompt template, or review note. A later complaint can then be discussed, but not reconstructed.

With RAIDT, the run-level evidence pack records the prompt, model identifier, retrieval snapshot, clause version, output, and reviewer sign-off. The mechanism that matters is preserved provenance: retrieval augmentation only improves governance when the retrieved material is captured and linked to the output. The score profile can then show whether Auditability and Traceability are weak, partial, or audit-ready using anchors 1=missing / 3=partial / 5=audit-ready. Mechanism-based explanation adds the reason this matters: the intervention is not valuable simply because it is technical, but because it creates a reviewable path from policy text to advice, which supports challenge, correction, and organisational learning.

• 12-RAIDT_DSR_Theory_M_v8
• 17-RAIDT-Sociotechnical_M_v6