• qa_deck_100#slide 71 · Prompting as a governance intervention

Role-based prompting is treated as a governance lever in RAIDT because it is not merely stylistic; it encodes a duty of care that changes what the model is permitted, expected, and forbidden to say. The prompt-design paper shows that role conditioning assigns an institutional persona, fixes vocabulary, tone, and escalation norms, and can require uncertainty language and explicit red-flag surfacing. In high-stakes settings such as clinical summarisation or credit decision support, those are governance choices because they shape what a decision-maker reads and how risks are disclosed.

Within the five pillars (Responsibility, Auditability, Interpretability, Dependability, Traceability), role prompts most directly strengthen Responsibility and Interpretability. They make scope, caution, neutrality, and audience alignment inspectable by reviewers, which is why the papers treat prompt design within the broader frame of influence methods as governance interventions rather than as ad hoc UX phrasing. Because RAIDT treats run as the unit of governance, a role prompt has to be versioned, linked to prompt IDs, and preserved with hashes, reviewer forms, and run logs in a run-level evidence pack. That documentation turns persona constraints into something that can be replayed and challenged.

The same papers also show the limit. Role-based prompting alone does not establish provenance, so its score profile is typically stronger on Responsibility and Interpretability than on Traceability. Using the anchors 1=missing / 3=partial / 5=audit-ready, role prompting often moves a weak baseline from missing governance cues to partial control, but it becomes audit-ready only when stacked with retrieval, stable logging, and other documented artefacts.

A bank uses an LLM to draft an adverse-action style explanation after a preliminary credit screening. A generic prompt may produce polished but unsafe text, for instance implying that a lending decision is final or speculating about reasons not supported by the record. A role-based prompt instead frames the model as a credit analyst writing for both a customer and an internal reviewer. It requires plain-language explanation, no speculative diagnosis of the applicant, an uncertainty statement where evidence is incomplete, and escalation to human review if the case involves thin-file data or conflicting records.

Under RAIDT, that role instruction is governance-relevant because the output can be tested against reviewer rubrics and preserved with the prompt version, output hash, and adjudication notes. The organisation can then show not only what the model said, but what institutional posture it was instructed to adopt at that run.

• 04-RAIDT_Prompt_Eng_V2
• 07-RAIDT_RLHF_V1