• workshop_dense_100#slide 45

In RAIDT, a prompt ID is the stable identifier for the prompt template or instruction asset, while the prompt version records the exact revision of that asset used in one run. Conceptually, they belong to configuration provenance inside the run-level evidence pack. The Evidence Review frames run-level evidence as a structured, versioned, reviewable object, and the Foundations paper specifies that organisations must be able to show what prompt and template were used in a particular run. A concise example would be PROMPT-ELIG-ADVICE v2.4: the ID names the governed prompt family, and the version points to the exact approved text active at the time.

This matters because RAIDT treats the run as the unit of governance. Governance questions are usually not abstract; they arise when an organisation must explain one output, one recommendation, or one disputed decision. Prompt IDs and versions allow reviewers to reconstruct that configured use event, relate the output to approved constraints, and distinguish prompt-driven changes from changes in retrieval, deployment, or oversight. They also support the five pillars (Responsibility, Auditability, Interpretability, Dependability, Traceability) by making prompt provenance explicit and reviewable. In practice, they feed the score profile because evidence completeness depends on stable identifiers, not recollection. If the run cannot show which prompt revision was active, the evidence pack becomes weaker for audit sampling, incident analysis, and organisational learning, and it is difficult to justify anchors 1=missing / 3=partial / 5=audit-ready in any high-stakes setting.

A public-service eligibility adviser uses a controlled prompt to draft guidance for a claimant. The run record shows PROMPT-ELIG-ADVICE v2.4, a model deployment ID, the retrieval snapshot ID for the policy text, and the reviewer?s final approval. Months later, the claimant challenges the advice. Because the prompt ID and version were preserved, the organisation can reconstruct the exact instruction template, check whether the approved rule wording required a cautious guidance-not-determination disclaimer, and assess whether the run complied with current controls at the time.

If the record contained only a generic note such as used eligibility prompt, the review would be much weaker. The organisation could not know which wording governed the output, whether the prompt had been superseded, or how to compare the run with other cases in the same score profile.

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