• qa_deck_100#slide 70 · Prompting as a governance intervention

Structured prompting changes the score profile unevenly rather than uniformly. In the prompt paper, instructional prompts produce a distinctly top-heavy profile: higher Interpretability because outputs follow a declared schema, higher Responsibility because uncertainty statements and red-flag surfacing appear more consistently, and a moderate Auditability gain when templates are versioned and logged as artefacts. Role-based prompts show a similar pattern, often with an extra lift in Responsibility because the audience and duty-of-care language are clearer. By contrast, zero-shot prompting yields the smallest and flattest profile, with more omissions, more drift, and weaker reviewability.

Using the RAIDT anchors 1=missing / 3=partial / 5=audit-ready, structured prompting most reliably moves runs upward on Interpretability and Responsibility, and often from missing toward partial on Auditability once prompt IDs, versions, and hashes are captured. Dependability improves, but only modestly: the schema reduces unwanted variation, yet the LoRA paper shows that adapter-based control is the stronger lever for stability and consistent structure. Traceability is the most constrained pillar. The RAG paper is clear that the largest Traceability gain comes from inspectable sources, citations, and retrieval logs, not from prompting alone. So the structured-prompt score profile is best understood as disciplined but incomplete: it strengthens the front-end governability of a run, while fuller audit-ready coverage across the five pillars usually requires stacking with LoRA and especially RAG.

Consider a finance support tool drafting a credit explanation. A structured prompt can force the output into fixed fields such as Decision Summary, Main Factors, Missing Information, Customer-Safe Explanation, and Uncertainty. Reviewers will usually score that run higher for Interpretability because the explanation is easier to read and compare, and higher for Responsibility because the model is less likely to hide uncertainty behind fluent prose.

However, the same run may still score only partial on Traceability if the explanation is not tied to policy excerpts or source records. If the bank later adds RAG over approved lending policy documents and logs the retrieved passages, the score profile changes again: Traceability and Auditability rise, while structured prompting continues to support clarity and consistency. That is the typical RAIDT pattern described across the papers.

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