• workshop_dense_100#slide 44

In RAIDT, a prompt registry is the governed record that stores prompt templates together with their identifiers, versions, status, owners, hashes, and change reasons, so prompting can be reviewed as part of organisational control rather than treated as informal craft knowledge. This definition fits the papers? wider claim that prompts must be versioned and logged, because prompting materially shapes outputs and therefore belongs inside configuration provenance. A registry is not the whole governance object by itself; rather, it supplies one of the key artefact layers that feed the run-level evidence pack when a specific use is examined. It helps establish which prompt template was active, what constraints it imposed, whether it was approved, and whether its integrity can be checked later.

A concrete RAIDT example would be a healthcare note-summarisation service using a structured discharge-summary template. The registry would hold the template identifier and version, the approved wording requiring an uncertainty statement, the responsible owner, the hash for integrity checking, and the recorded reason for any revision after policy review. When that prompt is linked to a particular run, reviewers can interpret the output in context and assess the score profile across the five pillars (Responsibility, Auditability, Interpretability, Dependability, Traceability). This is why the registry matters: RAIDT treats the run as the unit of governance, and prompting is one of the most consequential influence methods as governance interventions. A mature registry supports comparison across runs, change control over prompt evolution, and evidence-based scoring under the anchors 1=missing / 3=partial / 5=audit-ready. Without it, prompt text may exist, but governance remains too weak for reconstructable, contestable, and auditable GenAI use.

In a hospital discharge workflow, clinicians use a GenAI assistant to draft a summary from a note. The approved prompt requires fixed headings and an explicit uncertainty section. The prompt registry records that template as, for example, the current approved version for discharge summaries, names the clinical documentation lead as owner, stores the prompt hash, and notes that the last change was made to reduce overconfident wording.

When one summary is later reviewed after a complaint, the team can pull the linked run-level evidence pack and show not only the output, but also the exact prompt version and its governance status at the time of use. That makes it easier to explain why the output looked the way it did, whether the control wording was followed, and whether later prompt changes are relevant to the incident.

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