• qa_deck_100#slide 72 · Prompting as a governance intervention

Supervisors should read prompting inside the wider governance-intervention branch, not as an isolated usability device. In the prompting paper, a prompt becomes governance-relevant when it moves the model towards a safer or riskier posture, structures what end users read, and therefore enters the audit trail. Under RAIDT, prompting is best understood as one member of a family of influence methods as governance interventions. Its proper supervisory question is not simply whether the wording improved task performance, but whether the intervention changed the score profile across the five pillars (Responsibility, Auditability, Interpretability, Dependability, Traceability) and whether that change is evidenced at run level.

This wider reading matters because prompting has a characteristic governance signature. Instructional and role-based prompts often improve Responsibility and Interpretability by forcing clearer structure, uncertainty signalling, and audience-appropriate tone. However, the same papers also show that prompt-only control is fragile unless it is supported by a prompt registry, semantic versioning, output hashes, reviewer forms, and links to any retrieval context. Compared with LoRA, RAG, and RLHF, prompting is usually the easiest lever to deploy but the weakest stand-alone mechanism for Auditability and Traceability in high-stakes settings.

Accordingly, supervisors should treat the run as the unit of governance and examine the run-level evidence pack rather than the prompt text alone. The relevant reading is the per-run score profile, reviewer justification, and artefact completeness against the programme?s anchors 1=missing / 3=partial / 5=audit-ready. Where prompting lifts clarity but leaves provenance or lineage thin, supervisors should read it as a partial governance intervention that may need stacking with RAG or LoRA before it is suitable for stronger assurance claims.

In a hospital discharge-summary workflow, a supervisor should not approve an instructional prompt merely because it produces tidy headings such as symptoms, diagnosis, treatment, and red flags. The stronger question is whether that prompt version is registered, whether the output is hash-linked to the run, whether reviewers scored the result under RAIDT, and whether any retrieved guidelines are recorded.

If a zero-shot prompt yields fluent text but no source linkage, its score profile may still sit near the partial range on Auditability and Traceability even if Interpretability improves. By contrast, an instructional prompt paired with retrieval records and reviewer adjudication becomes a clearer governance intervention: the hospital can reconstruct the run, inspect omissions, and justify why the output was accepted or rejected.

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