• integrated_82#Q3.24

Stacked influence is the deliberate composition of more than one control surface around an LLM, typically prompt scaffolds plus one or more of LoRA/PEFT, RAG, and RLHF or DPO. In RAIDT terms, these are influence methods as governance interventions, not merely optimisation tricks. The papers describe the logic as modular and cumulative: prompts provide behavioural scaffolding and uncertainty discipline; LoRA provides stable, localised behavioural change with adapter lineage; RAG contributes provenance and contestable evidence; RLHF or DPO improves tone, refusal posture, and social alignment when its preference process is documented.

Stacked methods often score better because RAIDT rewards balanced governance coverage rather than isolated excellence. A single method may score strongly on one pillar yet remain weak elsewhere. Prompt-only configurations can be intelligible but brittle. PEFT can be stable and auditable yet still lack external grounding. RAG can make claims inspectable yet leave domain tone or institutional style under-specified. RLHF can improve Responsibility and user trust but introduces additional evidence complexity around annotators, reward models, and policy versions. When these methods are combined thoughtfully, the system achieves a fuller score profile across the five pillars (Responsibility, Auditability, Interpretability, Dependability, Traceability). The papers therefore treat stacked influence as the most governance-ready pattern for high-stakes work, while also warning that stronger performance comes with heavier documentation duties and more moving parts to monitor.

In public-service policy drafting, an analyst may need a short briefing note on a welfare-rule change. A prompt scaffold can force a standard format with uncertainty and escalation language. RAG can retrieve the current statute, guidance paragraphs, and implementation circulars. A LoRA adapter can keep the language aligned with departmental style. If citizen-facing tone matters, a documented RLHF or DPO layer can reduce unnecessarily harsh phrasing.

Such a stack usually scores better than a single control because the briefing becomes both readable and contestable. A reviewer can inspect the retrieved sources, verify the policy wording, and see whether the tone layer influenced the final phrasing. The result is not just a better answer; it is a better governed answer.

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