• workshop_dense_100#slide 21

In RAIDT, boundary conditions are the stated limits that define where the framework's constructs, mechanisms, and claims should be expected to hold most strongly. They matter because RAIDT is presented as a mechanism-based mid-range design theory, not a universal account of all AI governance. The papers identify several linked conditions. First, the technology condition: RAIDT is intended for generative AI whose outputs are materially shaped at run time by prompts, retrieved context, tools, adaptation layers, or alignment settings. Second, the organisational condition: it is intended for workflows where outputs influence records, decisions, services, or communications and where later reconstruction, review, or challenge is plausible. Third, the evidence condition: organisations must be able to preserve a bounded run-level evidence pack under suitable privacy, retention, and access controls. Fourth, the proportionality condition: evidence depth should vary with consequence and duty of care.

These conditions matter because they stop the framework from being overstated. RAIDT can show governance readiness through the five pillars (Responsibility, Auditability, Interpretability, Dependability, Traceability), a score profile, and anchors 1=missing / 3=partial / 5=audit-ready, but only when the evidence base exists. If run evidence is absent, RAIDT can reveal governance weakness; it cannot magically supply missing provenance. Boundary conditions therefore do analytical and practical work at once: they define the scope of valid claims, clarify implementation prerequisites, and make it possible to compare organisational uses without pretending that every GenAI task requires identical controls.

A healthcare provider uses GenAI to summarise consultation notes for chest-pain cases. This sits squarely within RAIDT's boundary conditions because the output enters the patient record, may influence follow-up care, and requires later review if something goes wrong. The provider can log the prompt constraints, model deployment, any retrieval used, the output, uncertainty statements, and the clinician's oversight decision. In that setting, RAIDT can be applied meaningfully because the run is material and the evidence can be preserved.

By contrast, a clinician's one-off personal brainstorm that is not saved, not relied upon, and not linked to patient care is much less suitable. The example shows why boundary conditions matter in RAIDT: they distinguish governable organisational use from low-significance use, and they tell the organisation when a full run-level evidence pack and scoring process are warranted.

• 08-RAIDT_Foundations_M_V50
• 11-RAIDT_Academic_Logic_M_v11
• 12-RAIDT_DSR_Theory_M_v8