• qa_deck_100#slide 69 · Prompting as a governance intervention

Structured prompting is a stronger governance intervention than ad hoc prompting because it converts the prompt from an improvised instruction into a governed artefact. In the prompt-design paper, the prompt is treated as a control surface that encodes audience, schema, constraints, abstention rules, and uncertainty language, then binds those choices to versioning, hashing, reviewer instruments, and change control. The RLHF paper makes the broader governance claim explicit: influence should be handled as design-time control surfaces rather than ad hoc developer conveniences. In RAIDT terms, structured prompting therefore does more than improve style. It creates inspectable evidence across the five pillars (Responsibility, Auditability, Interpretability, Dependability, Traceability).

Its governance advantage is clearest when run as the unit of governance. A structured prompt makes each run testable against declared requirements: Did the model follow the schema? Did it surface red flags? Did it state uncertainty when evidence was insufficient? Was the exact prompt version logged and replayable? Ad hoc prompting rarely provides that discipline. It drifts across users, leaves weak lineage, and makes post-hoc justification difficult even when the output sounds plausible. The papers also show the limit of this intervention: structured prompting usually improves Responsibility and Interpretability first, and can lift Auditability when prompt registries and hashes are present, but it still leaves Traceability partial and Dependability only moderate unless it is stacked with RAG or LoRA. That is why it is stronger than ad hoc prompting, but not sufficient as a complete governance stack.

In a clinical summarisation workflow, a structured prompt might require five fixed sections: Symptoms, Diagnosis/Impression, Treatment/Plan, Red Flags, and Uncertainty. It can also instruct the model to say explicitly when the note does not support a claim and to avoid speculative advice. The prompt is stored with a prompt_id, version, schema reference, and hash; the run also records model version, input hash, output hash, and reviewer scores.

That arrangement is materially stronger than an ad hoc instruction such as "summarise this consultation". With the structured version, reviewers can see whether the model omitted a red flag, whether uncertainty was signalled, and whether a prompt revision changed behaviour. If the organisation later needs to investigate an unsafe summary, the run-level evidence pack is reconstructable. With an ad hoc prompt, the output may still be readable, but the control logic, disclosure policy, and replay trail are much harder to verify.

• 04-RAIDT_Prompt_Eng_V2
• 07-RAIDT_RLHF_V1