S7.05 - Artefacts

S7.05 ? Artefacts

flowchart LR
    A1[High-level principles] --> B[RAIDT - run-level evidence framework]
    A2[Scattered logs] --> B
    A3[Isolated documents] --> B
    A4[Weak reconstructability] --> B

    B --> C[[Artefacts - designed governance objects]]
    C --> D[Run-level evidence pack]
    C --> E[Five-pillar score profile]
    C --> F[Reviewer reconstruction]
    C --> G[Policy alignment]
    C --> H[Organisational learning]
    C --> I[Evidence over assertion
Reviewability
Contestability
Audit readiness]

    J1[Healthcare discharge drafting] --> C
    J2[Financial review workflows] --> C
    J3[Public-service case handling] --> C
    J4[Prompt registry] --> C
    J5[Policy crosswalk] --> C
    J6[Governance dashboard] --> C

? Star S7 - Academic Theory and Design Logic

Star context: Positions RAIDT as a design-science, mechanism-based mid-range theory by showing which governance artefacts make responsible GenAI use observable, reviewable, and operational at the level of a single run.

Academic picture

Definition / background

Artefacts are designed objects that embody a governance logic in a form that can be used, inspected, and evaluated. In design science research, artefacts are the tangible outputs through which a theoretical contribution becomes practically actionable. They may be models, methods, constructs, procedures, templates, or instantiated systems. In RAIDT, the term is used in this design-science sense, but with a stronger governance emphasis.

Within GenAI governance, artefacts matter because organisational oversight cannot rely only on abstract principles such as fairness, transparency, or accountability. Governance needs objects that carry evidence, structure judgement, and support review. In RAIDT, those objects include the run-level evidence pack, the five-pillar scoring rubric, prompt registries, policy crosswalks, reviewer checklists, exception records, and traceable summaries of how a specific run was configured and assessed.

Artefacts are therefore different from raw logs, isolated screenshots, or ad hoc notes. Raw traces may contain useful data, but they are not yet governance artefacts unless they are organised into a form that supports interpretation, review, and decision-making. Likewise, a policy statement is not by itself a RAIDT artefact unless it is operationally linked to the evidence generated by a run. RAIDT belongs in this discussion because it is a framework that deliberately produces governance artefacts rather than leaving evidence assembly to chance.

This makes artefacts central to the relationship between run-level evidence, evidence packs, score profiles, and the five RAIDT pillars. The evidence pack is an artefact that consolidates proof. The scoring rubric is an artefact that structures evaluation. A policy crosswalk is an artefact that links evidence to organisational and regulatory expectations. Together, these artefacts allow a single run to be translated into a reviewable governance object.

Why this concept matters

Artefacts solve a practical governance problem: GenAI use is transient, context-sensitive, and often difficult to reconstruct after the event. Without well-designed artefacts, organisations may know that a model was used but still be unable to explain what happened in a specific run, what evidence supports the result, or whether the use was acceptable under policy. This produces a gap between governance rhetoric and operational proof.

The concept also avoids a recurring confusion. Many governance discussions treat documentation as an administrative afterthought. RAIDT treats artefacts as core design outputs. That distinction matters because the quality of governance depends on how evidence is structured, not only on whether data exists somewhere in the system. If artefacts are weak, inconsistent, or absent, reviewability and contestability collapse.

For organisations using GenAI, artefacts are the means by which principles become operational governance. They support internal review, external audit, model-risk conversations, incident response, and continuous improvement. They also help different audiences work from the same object: practitioners, managers, auditors, regulators, and researchers can all inspect the same run-level evidence pack rather than relying on inconsistent narratives.

Key idea: Artefacts matter because they convert fleeting GenAI activity into durable, reviewable governance objects that RAIDT can score, inspect, and use for organisational accountability.

What this item enables

It converts transient run events into durable governance objects that can be inspected after the run has finished.
It standardises how evidence, judgement, and policy alignment are recorded across different GenAI uses.
It connects technical traces, human review, and organisational rules in a single evaluable structure.
It enables the assembly of the run-level evidence pack and the justification of the five-pillar score profile.
It supports comparison, escalation, learning, and audit preparation across many runs over time.

Practical example / likely audience question

Audience question

Are artefacts in RAIDT just extra paperwork added after the real AI work has already happened?

Answer

The concern behind this question is that governance artefacts may look like bureaucratic overhead rather than a substantive part of responsible AI use. The direct answer is no: in RAIDT, artefacts are not merely paperwork added after the fact. They are the designed objects through which a run becomes governable.

A run may involve a model, a prompt, a user, contextual instructions, source material, an output, and a review decision. If those elements remain scattered across logs, screenshots, memory, and separate documents, the organisation has activity but not governance. RAIDT addresses this by creating artefacts that deliberately assemble these elements into a coherent review object. The run-level evidence pack is the clearest example because it gathers the relevant traces, decisions, and contextual metadata into one inspectable structure.

Consider a manager asking whether a problematic GenAI output can be reconstructed six weeks later. A generic AI governance approach may point to a policy document or a broad assurance statement, but that does not show what happened in the specific case. RAIDT handles the issue better because its artefacts are designed around the run itself. The reviewer can inspect the prompt used, the model version, any human approval step, the evidence attached, the pillar scores, and the basis on which the run was judged acceptable or contestable.

Practical example in RAIDT terms

In a healthcare setting, a hospital uses a GenAI assistant to draft discharge summaries from clinician notes and structured patient records. The specific run concerns a patient with multiple medications and a recent change in dosage.

The run-level governance issue is not simply whether the model can draft text. It is whether this particular discharge-summary run can be justified, reviewed, and corrected if a dosage instruction is incomplete or misleading. RAIDT would require artefacts that capture the prompt template, model version, source inputs available to the system, generated draft, clinician edits, final approval status, and any policy or safety checks applied.

The evidence needed would include a prompt registry entry, an output snapshot, user and reviewer identifiers, timestamps, policy crosswalk notes for clinical safety, and rubric-based scoring across the five pillars. Responsibility is affected because a clinician must remain accountable for sign-off. Auditability and Traceability are affected because the run must be reconstructable. Interpretability matters because reviewers need to understand how the output was framed and whether it can be explained. Dependability matters because discharge documentation must be reliable enough for clinical use.

The artefact layer improves governance readiness because the hospital can review the case as a complete governance object rather than as a scattered set of logs. If a concern arises, reviewers can reconstruct the run, identify where oversight succeeded or failed, and feed the lesson back into template design, workflow controls, and future RAIDT scoring.

Detailed link to RAIDT

Artefacts links to RAIDT in four ways.

First, RAIDT is a design-science contribution, and artefacts are the designed outputs through which the framework becomes usable rather than remaining only conceptual.
Second, because RAIDT treats the run as the unit of governance, artefacts capture the configuration, context, evidence, and review decisions associated with a specific run.
Third, artefacts populate the run-level evidence pack and provide the structured inputs needed to justify the five-pillar RAIDT score profile.
Fourth, artefacts support reviewability, contestability, audit readiness, and organisational learning because they preserve what happened, how it was judged, and what should improve next time.

Artefacts ? Run-level evidence ? Evidence pack ? RAIDT score profile ? Governance readiness

This chain matters because RAIDT does not treat governance as a static policy layer above AI use. It treats governance as something assembled and evidenced through artefacts that make each run available for inspection, comparison, challenge, and learning.

Link to the five RAIDT pillars

Artefacts affect all five pillars, but they are especially central to Auditability and Traceability because those pillars depend on durable, reconstructable governance objects.

Responsibility

Artefacts make responsibility visible by recording who initiated, reviewed, approved, or rejected a run and on what basis. They prevent accountability from dissolving into vague organisational ownership.