S4.05 - Prompt_registry

S4.05 ? Prompt registry

flowchart LR
    A1[Prompts treated as disposable text]
    A2[Unclear ownership and approval]
    A3[Hidden wording changes]
    A4[Weak reviewer reconstruction]
    B[RAIDT - run-level evidence framework]
    C[[Prompt registry - governed prompt artefact record]]
    D[Governance move - evidence over assertion]
    E[Run-level evidence pack]
    F[RAIDT score profile]
    G[Reviewer reconstruction]
    H[Organisational learning]
    I[Policy-aligned governance readiness]
    J1[Prompt ID and version]
    J2[Prompt hash]
    J3[Owner and approver]
    J4[Change rationale]
    J5[Task/domain label]
    J6[Linked run record]

    A1 --> B
    A2 --> B
    A3 --> B
    A4 --> B
    B --> C
    C --> D
    C --> E
    C --> F
    C --> G
    C --> H
    E --> I
    F --> I
    G --> I
    H --> I
    J1 --> C
    J2 --> C
    J3 --> C
    J4 --> C
    J5 --> C
    J6 --> C

? Star S4 - Evidence Architecture and Artefacts

Star context: Specifies the concrete fields and artefacts that make a run record inspectable. In RAIDT, the prompt registry is the governance layer that makes prompt use reviewable rather than informal, hidden, or anecdotal.

Academic picture

Definition / background

A prompt registry is the governed record through which an organisation defines, stores, maintains, and controls prompt templates used in generative AI work. In practical terms, it holds the canonical version of a prompt, who owns it, what status it has, how it changed over time, and why those changes were made. In RAIDT, this matters because prompt wording is not incidental. It can alter task framing, output style, acceptable evidence, risk exposure, and the degree to which a run can later be interpreted or challenged.

Conceptually, the prompt registry sits between informal prompt practice and formal run documentation. A saved prompt text on its own is useful, but it is not yet governance. Governance requires identity, version control, change rationale, accountable ownership, and the ability to connect a specific run back to the exact prompt artefact that shaped it. That is why RAIDT distinguishes the broader prompt registry from narrower items such as prompt ID and version, prompt hash, and run-level logging.

The item belongs inside RAIDT because RAIDT treats the run as the unit of governance. If a run is to be reviewable, the prompt used in that run must be reconstructable as a controlled artefact rather than remembered approximately. The prompt registry therefore supports the run-level evidence pack by linking prompt design decisions to actual system use, and it supports the five-pillar score profile by making those decisions assessable rather than assumed.

A prompt registry is also different from a generic prompt library. A library may support reuse and convenience. A registry supports evidence, accountability, and control. In RAIDT terms, it helps move prompt management from ad hoc craft practice to inspectable organisational governance.

Why this concept matters

Without a prompt registry, organisations often know that a model was used but cannot reliably show which prompt template governed the run, whether the wording had been changed recently, or whether the person using it departed from approved practice. This creates avoidable ambiguity in responsibility assignment, output interpretation, audit review, and post hoc learning.

The concept matters because prompts are a material part of system behaviour. If wording changes the scope of a task, the role assigned to the model, the instructions about evidence use, or the threshold for escalation, then prompt management is a governance issue rather than a convenience issue. A registry reduces the risk that meaningful behavioural change is hidden inside undocumented wording variation.

For organisations using GenAI, the registry helps separate three questions that are often confused: what prompt design was approved, what prompt version was actually used in a given run, and whether that prompt was suitable for the context. RAIDT needs all three distinctions because it aims to support reviewability, contestability, and continuous improvement at the level of concrete use.

Key idea: A prompt registry matters because RAIDT can only govern prompt-driven behaviour if prompts are treated as controlled artefacts linked to specific runs.

What this item controls

The canonical prompt templates that an organisation recognises as legitimate governance artefacts.
The identity and status of each prompt, including owner, approval state, and intended task domain.
Version history, including what changed, when it changed, and why the change was made.
Connections between prompt records and related evidence such as prompt IDs, hashes, task labels, tool use, and model settings.
The boundary between approved prompt design and informal prompt improvisation.
The conditions for reviewer reconstruction when a run needs to be examined, challenged, or compared.

Practical example / likely audience question

Audience question

Why is a prompt registry needed if RAIDT already stores the prompt text used in the run?

Answer

The concern behind this question is usually that saving the final text string appears sufficient for traceability. It is sufficient for partial reconstruction, but not for governance. A single saved text does not show whether that text came from an approved template, whether it was superseded, whether it had been modified by an operator, or why its wording had changed since earlier runs.

The direct answer is that a registry adds institutional meaning to prompt text. It turns a prompt from a raw input into a governed artefact with provenance, accountability, and change history. For example, a clinical summarisation team may save the exact prompt used in a run, but without a registry they may still be unable to show whether the wording that instructed the model to "prioritise likely diagnosis" was newly introduced, experimentally altered, or formally approved. That gap matters because reviewers need to know whether the behaviour arose from the model, the prompt design, or a departure from procedure.

RAIDT handles this better than generic AI governance because it ties the registry to run-level evidence rather than treating prompt management as a separate policy document. In RAIDT, the question is not merely whether prompts are documented somewhere. The question is whether a specific run can be connected to the exact governed prompt artefact that shaped that run and can therefore be reviewed on evidence.

Practical example in RAIDT terms

Consider a healthcare provider using a GenAI assistant to draft discharge-summary explanations for patients in plain English. One run produces a clinically acceptable output; another run, using the same model and task label, produces language that overstates certainty and omits a medication warning. The run-level issue is whether the difference arose from data, operator behaviour, decoding settings, or prompt wording.

A prompt registry allows reviewers to discover that the second run used a newly edited prompt template that instructed the model to "keep the message concise and confident" and that this change had been introduced to improve readability. The evidence needed includes the prompt ID, version, hash, owner, change rationale, approval status, timestamp of the revision, and the run record showing which version was invoked.

The RAIDT pillars most affected are Responsibility, Auditability, Interpretability, and Traceability, with Dependability also relevant because prompt drift can destabilise output quality. Governance readiness improves because the organisation can explain what changed, who authorised it, how it affected behaviour, and what corrective action is justified.

Detailed link to RAIDT

Prompt registry links to RAIDT in four ways.

First, it supports RAIDT's core idea that governance should attach to a concrete run rather than to abstract claims about safe or responsible AI use.
Second, it links the run to the governed prompt artefact that shaped the run's behaviour and therefore strengthens run-level evidence.
Third, it improves both the evidence pack and the score profile because reviewers can inspect prompt provenance, change control, and approval status rather than relying on unsupported assertions.
Fourth, it strengthens reviewability, contestability, audit readiness, and organisational learning by allowing prompt changes to be examined as explicit governance decisions.

Prompt registry -> Prompt identity and change control -> Run-level evidence -> Evidence pack -> RAIDT score profile -> Governance readiness

Link to the five RAIDT pillars

Responsibility

The prompt registry clarifies who is accountable for prompt design, revision, approval, and deployment. It prevents prompt wording from becoming an unowned source of behavioural change.