• workshop_dense_100#slide 68

In RAIDT, structured prompting means designing prompts with fixed schemas, explicit constraints, and mandatory uncertainty or abstention language so that outputs are easier to inspect, review, and govern. In the papers, this appears most clearly in the instructional and role-based prompt families, where prompts specify audience, disclosure rules, section ordering, red-flag requirements, and output-schema contracts. The key shift is that the prompt is no longer a disposable string; it becomes a versioned artefact linked to prompt registries, hashes, reviewer forms, and run logs.

Why this matters is straightforward. RAIDT treats influence methods as governance interventions, and structured prompting is the most accessible of those interventions because it can improve behaviour at inference time without retraining. It gives the organisation a stable way to shape explanation quality, uncertainty signalling, and safety language across runs. When run as the unit of governance, the structured prompt can sit inside a run-level evidence pack alongside the prompt version, input/output hashes, reviewer judgements, and any adapter or retrieval identifiers. That makes behaviour contestable rather than merely persuasive.

The papers also show why structured prompting matters but does not stand alone. It improves Interpretability and Responsibility first, can support Auditability when properly logged, but reaches a ceiling on Dependability and especially Traceability. LoRA contributes more stability; RAG contributes stronger provenance. Accordingly, structured prompting matters in RAIDT because it provides the disciplined surface on which stronger composite governance can be built.

A concrete RAIDT example is the clinical-scribe prompt described in the prompt paper: the model is asked to produce a structured summary with mandatory sections for Red Flags and Uncertainty, to keep rationale concise, and to stay within an approved clinical scope. The prompt record also links to a schema, a prompt hash, and the relevant run metadata.

This matters because a clinician or auditor can examine the exact control logic that shaped the answer. If the model misses a safety-critical red flag, the organisation can test whether the failure came from the prompt design, the model, or missing evidence. If the run is later stacked with LoRA for stable clinical register and RAG for guideline citations, the same structured prompt remains the visible governance surface that ties the whole run together.

• 04-RAIDT_Prompt_Eng_V2
• 06-RAIDT_RAG_V1