Q228 - Adapter_ID_PEFT_lineage_definition_example_and_why_it_matter

Q228 — Adapter ID / PEFT lineage — definition, example, and why it matters in RAIDT

← RAIDT · Star S4 - Evidence Architecture and Artefacts · primary item: S4.13 · Adapter ID / PEFT lineage

D. Evidence Architecture | Ordered by mind-map priority: inner circles first, then operational detail.

Appears in sources
Answer

In RAIDT, Adapter ID / PEFT lineage is the run-level record that identifies any LoRA or other PEFT component active in a run, together with the base model it modifies, the relevant training dataset reference, and a checkpoint hash or equivalent version marker. Conceptually, it sits inside configuration provenance within the run-level evidence pack. The foundations and evidence-review papers show why this field is needed: generative-AI governance cannot rely on model cards or programme-level audit routines alone, because behaviour and risk materialise in one configured use event. If PEFT changes what the system does, its identity and lineage must be captured as evidence rather than treated as an informal engineering detail.

Its importance in RAIDT is that it links run reconstruction to the five pillars (Responsibility, Auditability, Interpretability, Dependability, Traceability). Traceability is strengthened because reviewers can establish which adapted component contributed to the output. Auditability is strengthened because the run can be checked against approved versions and integrity markers. Dependability is strengthened because changes in stability can be associated with a specific adapter lineage rather than vaguely attributed to the model family. Responsibility is supported because approval, review, and escalation can be tied to a governed configuration. The result is a defensible score profile for a specific run. In short, RAIDT treats influence methods as governance interventions, so adapter lineage matters whenever PEFT materially shapes organisational action.

Practical example

A concrete RAIDT example is an HR workflow in which a manager uses a GenAI assistant to draft a shortlist justification. The organisation records the prompt template version, the model deployment, any retrieved policy text, the output hash, and the Adapter ID / PEFT lineage for the recruitment-specific LoRA component.

That single field allows a later reviewer to see not just that ?an LLM? was used, but which adapted checkpoint shaped the recommendation, what base model it sat on, and which dataset reference belongs to the adaptation lineage. If the employee contests the shortlist, the organisation can reconstruct the governed run, assess whether the authorised adapter was used, and judge the evidence against the anchors 1=missing / 3=partial / 5=audit-ready. In RAIDT terms, that is why Adapter ID / PEFT lineage is not optional metadata; it is part of what makes the run reviewable at all.

Sources in RAIDT papers
Powered by Forestry.md