• qa_deck_100#slide 45 · Retrieval, tools, adapters, and alignment evidence

Adapter ID and PEFT lineage must be recorded because RAIDT treats the run as the unit of governance, and a run is materially shaped not only by the base model but also by any active adaptation layer. The papers are explicit that generative-AI behaviour depends on run-time configuration, including prompt structure, retrieval context, tool use, settings, and parameter-efficient adaptation such as LoRA. In that logic, an adapter is not a minor implementation detail: it is part of the operative configuration provenance that determines what the system actually did in one governed use. Recording the adapter ID, linked base model, training dataset reference, and checkpoint hash therefore makes the run-level evidence pack reconstructable rather than merely descriptive.

This matters across the five pillars (Responsibility, Auditability, Interpretability, Dependability, Traceability). For Traceability and Auditability, adapter lineage shows which component contributed to the output and whether the run can be reconstructed months later in a dispute. For Dependability, the foundations paper notes that domain adaptation can improve task consistency but also creates version-control and change-management obligations; without lineage, instability or drift cannot be attributed to a specific configuration. For Responsibility, lineage clarifies which governed component was approved, deployed, or escalated. In practice, a score profile cannot be defended if a materially influential adapter is invisible. Missing lineage would push the record towards the lower anchors 1=missing / 3=partial / 5=audit-ready because reviewers could not verify which adapted behaviour was actually in force.

Consider the HR shortlist justification scenario used in the RAIDT papers. An organisation uses a base language model with a LoRA adapter tuned for recruitment-style reasoning and stores the generated shortlist note in a personnel file. Six months later, an unsuccessful applicant challenges the process and asks what system configuration informed the recommendation.

If the run-level evidence pack contains the adapter ID, the base model identifier, the dataset reference attached to that adapter, and the checkpoint hash, reviewers can reconstruct whether the approved recruitment adapter was used, whether it matched the authorised version, and whether the output came from the governed configuration rather than from an older or experimental checkpoint. That makes the case reviewable by HR, compliance, and internal audit. If those fields are absent, the organisation may know the model family in general terms but still be unable to show what adapted behaviour shaped this contested run.

• 08-RAIDT_Foundations_M_V50
• 18-RAIDT-Technical-Foundation_M_v04