• workshop_dense_100#slide 66

A governance intervention in RAIDT is an intervention whose purpose is not merely to optimise output, but to make that output governable. The decisive feature is evidential as well as behavioural. If a method changes what the model says, how uncertainty is expressed, which sources are attached, or how a reviewer can reconstruct the run, it is operating as a governance intervention. The prompting paper makes this explicit for prompt design; the LoRA and RAG papers extend the same logic to modular adapters and provenance-first retrieval. In all three cases, the intervention becomes part of the technical record rather than remaining tacit developer craft.

A simple example is a clinical summarisation pipeline that uses an instructional prompt, a domain LoRA adapter, and RAG over approved guidance. The prompt structures mandatory sections and uncertainty statements; the adapter stabilises specialty tone; the retriever links claims to inspected sources. What matters in RAIDT is that these levers can be tied to a run-level evidence pack containing prompt version, adapter lineage, retrieval context, output hashes, and reviewer notes. The resulting score profile is then judged across the five pillars (Responsibility, Auditability, Interpretability, Dependability, Traceability) using the anchors 1=missing / 3=partial / 5=audit-ready.

This matters because RAIDT rejects the view that governance is external to system design. The papers show that governance quality depends on how influence is engineered. A prompt may improve readability, but without logging and source linkage it offers weak assurance. By contrast, an intervention designed for evidence capture supports supervision, contestability, corrective action, and standards mapping. In RAIDT terms, governance interventions convert abstract trust claims into run-level artefacts that auditors, supervisors, and affected stakeholders can actually inspect.

A public-service team drafting a briefing on local health provision could use a prompt-only workflow that summarises documents into a short note. That may be efficient, but if a councillor later challenges a claim, the team may not be able to show which version of the prompt produced the statement or which documents supported it.

If the same workflow is rebuilt as a governance intervention, the briefing run stores the prompt ID, the approved LoRA adapter version for institutional tone, the retrieved policy documents, the output hash, and the reviewer form. The briefing is then easier to contest, correct, and defend. The intervention has changed both behaviour and the evidence available for oversight.

• 04-RAIDT_Prompt_Eng_V2
• 05-RAIDT_LoRA_V2
• 06-RAIDT_RAG_V1