• qa_deck_100#slide 67 · Prompting as a governance intervention

RAIDT treats prompting as a governance intervention because a prompt is not merely a convenience string; it is a control surface that changes what a model says, how it says it, and what a user is likely to rely upon. In the papers, prompts encode task boundaries, output contracts, uncertainty language, red-flag surfacing, tone, and audience duties. In healthcare, finance, policy, and cybersecurity settings, those features shape consequential reading and action. For that reason, prompting is framed within the five pillars (Responsibility, Auditability, Interpretability, Dependability, Traceability), rather than as a narrow optimisation trick. A prompt that can move a system towards a safer or riskier posture is, in RAIDT terms, part of governance by design.

The governance consequence is that prompts must leave evidence. The papers consistently argue that prompt drift, undocumented edits, and unlinked outputs undermine record-keeping, transparency, and human oversight. Hence influence methods as governance interventions must be logged, versioned, hashed, and tied to reviewer judgement. In practice, the run as the unit of governance means that each run should preserve the prompt identifier and version, model/version metadata, hashes, adjudication notes, and any retrieval or adapter links. That bundle functions as a run-level evidence pack, allowing auditors to reconstruct why an output appeared in a given form. RAIDT therefore treats prompting as compliance-relevant because it affects behaviour ex ante and because its effects must be demonstrable ex post.

In a clinical summarisation workflow, a hospital may use an instructional prompt that requires four bullets: symptoms, diagnosis, treatment, and red flags, plus an explicit statement when evidence is insufficient. That prompt is not just formatting advice. It shapes whether urgent warnings are surfaced, whether uncertainty is visible to the clinician, and whether the output is safe to act upon.

Under RAIDT, the organisation would register that prompt, assign a version, log the run ID, keep input and output hashes, and retain reviewer comments on whether the summary met clinical and governance expectations. If a later review finds that a red flag was omitted, the team can trace the exact prompt version and compare it with subsequent revisions. The prompt is therefore governed as an intervention in decision support, not treated as disposable wording.

• 04-RAIDT_Prompt_Eng_V2
• 07-RAIDT_RLHF_V1