• qa_deck_100#slide 37 · Minimum run identity and configuration record

Every RAIDT record should begin with a Run ID because RAIDT defines the run as the unit of governance, not the model, system, or policy document in the abstract. The papers consistently argue that governance becomes meaningful only when one configured use can be reconstructed, reviewed, and contested in context. A Run ID provides the stable reference that binds the run-level evidence pack into one bounded evidential object rather than a loose collection of prompts, logs, hashes, outputs, and review notes. Without that identifier, evidence fragments may exist, but they are harder to assemble into an auditable account of one material use.

This is especially important for Auditability and Traceability within the five pillars (Responsibility, Auditability, Interpretability, Dependability, Traceability). The foundations paper gives Auditability anchors such as Run ID/time, prompt and version, model deployment ID, parameters, output hash, and check logs. In practice, the Run ID is the primary key that lets reviewers connect those fields, verify integrity, and distinguish one run from another when similar prompts or models are reused. It is also what makes the score profile defensible, because evidence pointers must refer to one specific governed event rather than to general system behaviour.

The Run ID also matters organisationally. RAIDT expects run records to link into risk registers, incident tickets, audit reports, and change-management logs. That linking function is crucial when influence methods as governance interventions alter behaviour at run time through retrieval, prompting, adapters, or alignment policies. In that setting, the Run ID is not clerical metadata; it is the spine that makes reconstruction, sampling, comparison, and later challenge possible. This is why RAIDT treats a stable run identifier as foundational to a run-level evidence pack and to anchors 1=missing / 3=partial / 5=audit-ready.

Consider a public-service eligibility workflow in which a caseworker uses GenAI to draft advice on whether an applicant meets a benefit rule. Months later, the applicant contests the decision. If the record begins with a Run ID, reviewers can pull the exact prompt template version, the model deployment ID, the retrieval snapshot containing the policy clause used, the output hash, and the human review note tied to that one event. The audit can test whether the cited rule text was the correct version and whether oversight occurred before the advice was issued.

If there is no Run ID, the organisation may still possess logs, prompts, or documents, but it cannot confidently show which pieces belong together. Similar runs from the same week may be confused, source snapshots may be misattributed, and the resulting score profile will be weak because the evidence pack is no longer securely linked to one governed run.

• 08-RAIDT_Foundations_M_V50
• 13-RAIDT-Evidence-Review_M_v10