• workshop_dense_100#slide 67

Prompting is the practice of shaping model outputs through instructions, constraints, roles, schemas, and related textual cues. In RAIDT, however, that basic definition is extended: prompts are treated as governed artefacts because they shape behaviour without changing model weights and because they influence what people read, trust, and act upon. The prompt paper shows that instructional and role-based prompts can improve clarity, uncertainty signalling, and domain-appropriate tone, especially in high-stakes tasks such as clinical summarisation, credit explanation, and policy briefing. RAIDT therefore locates prompting within the five pillars (Responsibility, Auditability, Interpretability, Dependability, Traceability), asking not only whether a prompt improves output quality, but whether it produces inspectable, replayable, policy-relevant evidence.

Why it matters follows directly from that framing. Prompting is often the first and cheapest control lever, but RAIDT refuses to leave it at the level of tacit craft. A prompt should sit inside a registry, be versioned, and be tied to hashes, reviewer forms, and any provenance artefacts produced during a run. That is why a score profile matters: prompting can lift Interpretability and Responsibility quickly, but Auditability, Dependability, and Traceability depend on whether the prompt is documented and whether the surrounding evidence reaches anchors 1=missing / 3=partial / 5=audit-ready. Under run as the unit of governance, the prompt becomes part of a run-level evidence pack rather than an invisible instruction string.

A concrete example is de-identified clinical note summarisation. A role-based prompt might instruct the model to act as a senior clinical scribe, produce a four-part summary, surface red flags, and state explicitly when evidence is insufficient. This usually improves readability for clinicians and makes the rationale more audience-appropriate.

In RAIDT, the organisation would not stop at observing that the answer looks better. It would preserve the prompt version, model version, input and output hashes, reviewer scores, and, where available, retrieval identifiers for the evidence shown to the user. If repeated runs remain stable and the claims are source-linked, the prompt contributes to a stronger score profile across the five pillars. If those artefacts are absent, the same prompt remains useful operationally but weak as governance evidence.

• 04-RAIDT_Prompt_Eng_V2
• 06-RAIDT_RAG_V1