• integrated_82#Q3.22

Within RAIDT, PEFT/LoRA is best understood as one of the influence methods as governance interventions: it alters model behaviour through a small adapter while leaving the foundation model unchanged. The LoRA paper frames this as parameter-efficient control, not merely optimisation. Governance value comes from modularity. Because the adapter is a discrete delta, organisations can register its scope, data slice, target modules, hyperparameters, evaluation results, and risk notes in an adapter card, then connect that record to prompt registries and hash-linked run logs. This is a stronger intervention than informal prompting because the behavioural change is inspectable, attributable, and revocable.

It can improve dependability because the frozen base model narrows the blast radius of change and the papers report lower output variance and more stable structure under repeated runs than prompt-only baselines. It can improve version control because every substantive change creates a new adapter identity and lineage trail rather than silently overwriting a full model. In RAIDT terms, that means better evidence across the five pillars (Responsibility, Auditability, Interpretability, Dependability, Traceability), even though Traceability still benefits from RAG when claims must be source-anchored. The practical governance question is therefore not simply whether performance improved, but whether a given run can be replayed, scored, and challenged. PEFT/LoRA helps answer yes by linking the score profile for a run to named artefacts, reviewer judgements, and rollback procedures.

In healthcare, a trust deploying clinical note summarisation can use a LoRA adapter specialised for discharge-summary structure and red-flag surfacing while keeping the underlying model fixed. The governance intervention is not only the adapter itself, but the package around it: adapter card, dataset notes, prompt registry, reviewer rubric, consistency checks, and run logs. If a later release begins to overstate diagnoses or becomes too rigid outside its specialty, the team can inspect the adapter lineage, compare the score profile across versions, and revert to the previous adapter without retraining the whole model. Under anchors 1=missing / 3=partial / 5=audit-ready, that moves the deployment from ad hoc control towards auditable change control and more dependable clinical outputs.

• 05-RAIDT_LoRA_V2
• 04-RAIDT_Prompt_Eng_V2