S3.02 - Evidence_object

S3.02 ? Evidence object

flowchart LR
    A[Fragmented evidence landscape
logs, policy, screenshots, recollections] --> B[RAIDT
run-level evidence framework]
    H[Practical run artefacts
prompt, inputs, outputs, metadata, review notes] --> C[[Evidence object
bounded proof artefact for one run]]
    B --> C
    C --> D[Evidence pack]
    C --> E[RAIDT score profile]
    C --> F[Reviewer reconstruction
and contestability]
    D --> G[Governance readiness
and organisational learning]
    E --> G

? Star S3 - Run-Level Evidence Logic

Star context: Explains how RAIDT turns run-level evidence into a bounded governance artefact that can be inspected, reconstructed, compared, and challenged rather than treated as an abstract claim of assurance.

Academic picture

Definition / background

An evidence object in RAIDT is a bounded set of records assembled around one governed run so that a governance claim can be supported, inspected, compared, or challenged. The phrase matters because RAIDT does not treat evidence as a vague collection of anything that might later prove useful. Instead, it treats evidence as an identifiable object with scope, boundaries, contents, and a governance purpose.

Conceptually, the evidence object sits between raw traces and formal judgement. It is more structured than a loose archive of files, screenshots, and system logs, because it defines what belongs to the governed run and why. It is also more operational than a general assurance statement, because it contains the concrete artefacts from which reconstruction and review can actually occur. In that sense, the evidence object is the practical governance artefact that makes a run examinable.

This matters in generative AI governance because many existing frameworks speak about transparency, assurance, or documentation without clearly specifying the object that reviewers inspect when they test a real use event. RAIDT resolves that ambiguity by defining the run-level evidence object. The object can contain prompt material, inputs, outputs, metadata, human review notes, workflow decisions, and contextual markers, but its value lies in being bounded and inspectable rather than merely extensive.

Within RAIDT, the evidence object belongs directly to the logic of run-level evidence. It is the artefact from which a run-level evidence pack can be assembled and through which a five-pillar score profile can be justified. Without an evidence object, scoring becomes weakly grounded, comparisons across runs become inconsistent, and challenges to governance claims are harder to answer. The evidence object therefore links the run, the evidence pack, the score profile, and the practical goal of governance readiness.

Why this concept matters

The evidence object solves a persistent governance problem: organisations often have fragments of information about GenAI use, but not a defined object of proof that can be reviewed after the event. When an output is disputed, a decision is questioned, or a supervisor asks how one run should be assessed, the organisation needs more than policy and more than technical telemetry. It needs a bounded evidential unit that tells reviewers what they are inspecting.

The concept also avoids confusion between evidence collection and evidence governance. Collecting data is not the same as defining the artefact through which a claim can be examined. By naming the evidence object, RAIDT clarifies what must be assembled, retained, and reviewed in order to support reconstructability, comparability, and challenge.

If the evidence object is missing, governance becomes fragile. Reviewers may rely on partial logs, anecdotal recollection, or selectively retrieved records. Auditability weakens because no shared object exists for inspection. Traceability weakens because links between prompt, output, actor, and review step can break. Responsibility weakens because decision points and human interventions may not be evident. RAIDT uses the evidence object to prevent governance from collapsing into scattered documentation and post hoc interpretation.

Key idea: The evidence object matters because RAIDT needs a bounded, inspectable proof artefact for each governed run, not just a loose collection of records or high-level assurance claims.

What this item captures

The bounded evidential scope of one governed GenAI run.
The records that are necessary to support or challenge a claim about that run.
The connection between prompts, inputs, outputs, settings, human actions, and review decisions.
The minimum structure needed for reconstruction, comparison, and contestability.
The practical object from which an evidence pack can be assembled.
The evidential basis for justifying RAIDT pillar scores with something reviewable rather than impressionistic.
The shift from scattered documentation to a governance artefact that can be inspected by supervisors, auditors, reviewers, or practitioners.

Practical example / likely audience question

Audience question

What object does RAIDT actually add that generic AI governance frameworks often leave undefined?

Answer

The concern behind this question is that many governance frameworks already ask for documentation, logging, or assurance, so it can seem unnecessary to introduce another term. The direct answer is that RAIDT adds a run-level evidence object: a bounded governance artefact containing the records needed to inspect one specific GenAI run.

That distinction matters because generic approaches often identify types of documentation but do not define the concrete object a reviewer examines when testing a claim about a real use event. In practice, this leaves organisations with fragments such as model documentation, workflow notes, screenshots, or system logs, but no agreed evidential unit. RAIDT closes that gap by saying that governance attaches to the run and that the run is accompanied by an evidence object that can be inspected and challenged.

A practical example is an enterprise team using GenAI to draft a client-facing risk summary. If concerns later arise about the accuracy of a claim in the draft, generic governance may point to policy, staff training, or vendor assurances. RAIDT goes further by asking for the evidence object for that run: the prompt, source material, model settings, output draft, human edits, approval notes, and contextual metadata. That gives reviewers something concrete to inspect. RAIDT therefore handles the issue better than a generic approach because it defines the inspectable object of governance rather than assuming evidence will somehow be available when needed.

Practical example in RAIDT terms

Consider a finance setting in which an analyst uses a GenAI assistant to draft a preliminary anti-money-laundering case summary for internal review. The use case is not the final compliance decision; it is a productivity-support run that still carries governance risk because the generated summary may shape how a human reviewer interprets transactions and flags concerns.

The run-level issue is whether the summary can later be justified if a case is escalated or criticised. The evidence needed includes the task purpose, the analyst's role, the source transaction data supplied to the model, the prompt template, any instructions about risk thresholds, the model version or tool configuration, the generated text, the analyst's corrections, the compliance review note, and the final disposition of the summary.

In RAIDT terms, those records become the evidence object for that run. Responsibility is affected because the organisation must show who initiated and checked the summary. Auditability is affected because a reviewer must be able to inspect what was generated and how it was handled. Interpretability is affected because the prompt and review notes help explain why the output took the form it did. Dependability is affected because repeated cases can be compared for consistency and failure patterns. Traceability is affected because the artefacts are tied to time, actor, source material, and downstream review. The evidence object therefore improves governance readiness by turning a potentially disputed AI-assisted action into an examinable case rather than an unstructured memory of what happened.

Detailed link to RAIDT

Evidence object links to RAIDT in four ways.

First, it gives practical form to RAIDT's core commitment to evidence over assertion by defining the actual artefact through which governance claims are tested.

Second, it links directly to the run because RAIDT treats the run as the unit of governance and the evidence object as the bounded proof artefact attached to that unit.

Third, it supports the evidence pack and the score profile because the evidence object provides the inspectable material from which packaging, review, and five-pillar scoring can be justified.

Fourth, it strengthens reviewability, contestability, audit readiness, and organisational learning because reviewers can inspect the same object when reconstructing a disputed or exemplary run.

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

Link to the five RAIDT pillars

Responsibility

The evidence object supports Responsibility by showing who was involved in the run, what role they held, and which review or approval actions they performed.