• qa_deck_100#slide 77 · RAG, PEFT, RLHF-type controls, and stacked configuration

Stacked configurations usually score better than single controls because RAIDT evaluates a governance posture rather than a single technical trick. In the papers, prompting, LoRA/PEFT, RAG, and RLHF/DPO each strengthen different parts of the score profile across the five pillars (Responsibility, Auditability, Interpretability, Dependability, Traceability). Prompt structure improves Interpretability and often Responsibility by enforcing schema, uncertainty language, and audience-aware tone. LoRA improves Dependability and Auditability because adapter deltas are compact, versionable, and easier to roll back. RAG contributes the strongest gains in Traceability and Interpretability because claims can be tied to inspectable sources. RLHF can improve Responsibility by sharpening tone, refusal behaviour, and red-flag surfacing, although it also creates new provenance obligations.

The important point is complementarity. A single method leaves predictable gaps: prompt-only runs remain weak on provenance, LoRA alone does not fully ground claims, RAG alone does not stabilise domain idiom as strongly as PEFT, and RLHF alone can weaken Auditability if reward lineage is opaque. Stacked influence therefore works because it distributes labour across influence methods as governance interventions. The papers repeatedly show the resulting configuration broadens the score profile, raises composite RAIDT outcomes, and in some programme settings reaches the strongest audit-ready pattern only when PEFT and RAG are combined, with RLHF added where tone and safety need to be governed. In short, stacked methods score better because no single lever satisfies all evidential and behavioural requirements at once.

In finance, a bank generating adverse-action style explanations can start with a structured prompt that forces plain-language reasoning and uncertainty disclosure. A LoRA adapter then stabilises the institution?s credit vocabulary and formatting, while RAG retrieves the relevant lending-policy passages and product rules. If the bank also wants a more careful customer-facing tone, it may add RLHF with documented preference guidelines.

This stack scores better than a prompt-only control because the output is not merely fluent. It is source-anchored, stylistically stable, easier to replay, and easier to contest. Reviewers can see which prompt version ran, which adapter was loaded, which policy snippets were retrieved, and whether a tone layer shaped the wording. That richer evidence chain improves both the practical decision workflow and the RAIDT assessment.

• 05-RAIDT_LoRA_V2
• 06-RAIDT_RAG_V1
• 07-RAIDT_RLHF_V1