• integrated_82#Q1.12

The boundaries and limitations of RAIDT are defined in a disciplined way. First, RAIDT is bounded to generative AI systems whose behaviour is materially shaped at run time by prompts, retrieved context, tool use, adaptation layers, or alignment settings. It is not proposed as a general theory for every algorithmic system. Secondly, it is bounded to organisational settings in which outputs influence decisions, records, communications, or services, and where later reconstruction, review, or challenge is plausibly required. This is why RAIDT repeatedly frames the run as the unit of governance and centres governance on the run-level evidence pack rather than on model cards or abstract policy claims.

Thirdly, RAIDT assumes evidential feasibility. The framework works best where organisations can capture bounded evidence under suitable retention, privacy, and access controls. Where evidence cannot be stored consistently, RAIDT can expose the governance gap, but it cannot resolve it. Fourthly, RAIDT is explicitly proportionate: not every use needs the same instrumentation depth, so low-consequence drafting may justify a lighter evidence pack than eligibility advice, incident response, or clinical support. Finally, the framework is limited by trade-offs and non-universality. The score profile across the five pillars (Responsibility, Auditability, Interpretability, Dependability, Traceability) makes those trade-offs visible, especially when influence methods as governance interventions improve one pillar while weakening another. The anchors 1=missing / 3=partial / 5=audit-ready therefore delimit evidence quality for a given run, not total assurance for every domain, workflow, or legal regime.

A public-service agency using GenAI to draft eligibility advice illustrates these boundaries clearly. RAIDT is highly relevant if staff must later show which policy clause, retrieval snapshot, prompt template, and reviewer decision shaped one disputed explanation. In that case, the run-level evidence pack supports contestability and audit.

However, RAIDT is less useful for trivial, ephemeral drafting with no governance significance, and it is weakened if privacy law or technical architecture prevents preservation of the retrieved policy text or reviewer trail. The framework can still indicate that auditability and traceability are weak, but it cannot create missing evidence retrospectively. This matters because organisations should treat RAIDT as a bounded governance method for material runs, not as a universal substitute for service law, policy interpretation, or administrative judgement.

• 08-RAIDT_Foundations_M_V50
• 12-RAIDT_DSR_Theory_M_v8