• workshop_dense_100#slide 73

A stacked configuration is a governed combination of prompt design, grounding, adaptation, and sometimes alignment, assembled so that different methods cover different RAIDT needs. A typical pattern in the papers is prompt schema plus LoRA/PEFT plus RAG, with optional RLHF when tone or refusal behaviour is mission-critical. The reason it matters is straightforward: RAIDT does not ask only whether a model can answer, but whether the answer can be owned, checked, interpreted, reproduced, and traced. Stacking improves that governance fit because each layer contributes a distinct control function.

This matters especially in high-risk settings where the run, not the abstract model, must be defended. For that reason, organisations should treat the run as the unit of governance and interpret the resulting score profile as a property of the whole configuration rather than of one component in isolation. Prompting contributes structure and explicit uncertainty. LoRA contributes stable domain idiom and adapter lineage. RAG contributes citations, corpus hashes, and contestability. If alignment is added, it may improve Responsibility but also raises evidence complexity. The papers consistently show that the strongest deployments are those where the stack is paired with registries, hashes, reviewer forms, and replayable logs. Stacked configuration therefore matters in RAIDT because it turns dispersed controls into a governable system rather than a fluent but weakly evidenced output pipeline.

In education, a university using an LLM for essay feedback could run a stack with a rubric-driven prompt, a small LoRA adapter tuned to institutional marking language, and RAG over the approved rubric and exemplar comments. The generated feedback can then cite rubric items and relevant text spans rather than offering generic advice.

That matters for RAIDT because assessors can review not only the wording but the evidence behind it. If a student contests the feedback, the university can inspect the prompt version, the rubric passages retrieved, and the adapter used for style consistency. The stack therefore supports fairness, contestability, and dependable formatting in a way that prompt-only feedback usually cannot sustain in high-stakes assessment.

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