• workshop_table17#tag-band S6 · 95–120 min

In RAIDT, prompting, RAG, LoRA/PEFT and RLHF/DPO do not constitute the project core because the core object is the governance artefact produced when organisations treat the run as the unit of governance. Across the papers, these methods are treated as influence methods as governance interventions: they shape run-time behaviour and alter what evidence can be captured, reviewed and contested. Their role is therefore instrumental and evidential rather than foundational. Prompting can support Interpretability by enforcing structure, uncertainty statements and constrained rationale. RAG can support Traceability and Auditability when retrieval snapshots, identifiers and hashes are preserved. LoRA/PEFT can support Dependability and Traceability when adapter lineage is versioned and linked to the base model. RLHF/DPO fit the papers' broader category of preference-based alignment, which can strengthen Responsibility when the active policy layer and refusal logic are recorded as part of the run.

RAIDT remains centred on the run-level evidence pack, the five pillars (Responsibility, Auditability, Interpretability, Dependability, Traceability), and the score profile rather than on any single optimisation method. This is why the same technique may support governance in one run yet weaken it in another: a RAG configuration without preserved retrieval snapshots is not strongly traceable, and alignment without logged policy metadata can reduce reviewability. The scoring logic makes this explicit through anchors 1=missing / 3=partial / 5=audit-ready, applied to evidence rather than to vendor claims or benchmark performance. The component papers therefore support RAIDT by showing how technical configuration choices become governable only when they are bounded inside reconstructable run records, so that organisations can compare interventions across contexts without mistaking the interventions for the governance framework itself.

Consider a hospital discharge-summary assistant. The team uses a structured prompt that forces headings and an Uncertainty/limits section, adds RAG over internal clinical guidance, and deploys a LoRA adapter tuned for local documentation style. A preference-alignment layer blocks unsupported medication advice. These components may improve the service, but under RAIDT they matter chiefly because each run-level evidence pack records the prompt version, retrieval snapshot identifiers and hashes, base-model and adapter identifiers, active alignment policy, output hash, and clinician sign-off.

Governance then turns on the score profile across the five pillars (Responsibility, Auditability, Interpretability, Dependability, Traceability), not on the mere presence of RAG or PEFT. A run with source citations but no preserved retrieval snapshot may still sit only at the anchors 1=missing / 3=partial / 5=audit-ready for Traceability. By contrast, a fully logged run can be reconstructed months later if a patient or auditor challenges the summary. That is how the methods support RAIDT without becoming the core project.

• 08-RAIDT_Foundations_M_V50
• 11-RAIDT_Academic_Logic_M_v11
• 12-RAIDT_DSR_Theory_M_v8