C0.04 - Evidence_pack

C0.04 ? Evidence pack

flowchart LR
    A[Fragmented evidence
logs, prompts, approvals, policy references] --> B[RAIDT
run-level evidence framework]
    H[Practical evidence elements
run ID, task, prompts, settings, outputs, review, retention] --> C[[Evidence pack
structured record of one run]]
    B --> C
    C --> D[Reviewer reconstruction]
    C --> E[RAIDT score profile]
    C --> F[Contestability and audit readiness]
    E --> G[Governance readiness and organisational learning]

? Star C0 - RAIDT Core, Definition, Values, Claims and Innovation

Star context: Defines the project identity of RAIDT by showing that the framework's first practical governance artefact is a structured evidence object for one run, not merely a policy statement or assurance claim.

Definition / background

An evidence pack is the structured, review-ready record of one RAIDT run. It assembles the run-level evidence associated with a specific use of a generative AI system and presents it in a form that can be inspected by a supervisor, auditor, manager, regulator, complaint handler, or researcher. In practical terms, the pack links identifiers, task context, prompts, configuration, retrieval or source material, model settings, outputs, integrity checks, human review actions, decisions, and retention rules.

This distinction matters conceptually. Run-level evidence is the underlying evidential material produced by or around one run. The evidence pack is the organised governance object created from that material. In other words, run-level evidence is the substance, while the evidence pack is the structured presentation of that substance for review and judgement. The pack is therefore more than a log bundle and more governance-oriented than a raw technical trace.

Within RAIDT, the evidence pack is one of the framework's two practical outputs, alongside the five-pillar score profile. The pack supports the score profile by giving reviewers a documented basis for judging Responsibility, Auditability, Interpretability, Dependability, and Traceability. Without an evidence pack, a score profile risks becoming insufficiently justified; without run-level evidence, an evidence pack cannot be meaningfully assembled.

This item belongs inside RAIDT Core because it shows how the framework turns the abstract commitment to evidence over assertion into an operational artefact. RAIDT is not only a way of saying that governance should be evidence-based. It specifies what that evidence should look like when organised for real organisational scrutiny.

Why this concept matters

The evidence pack solves a practical governance problem: even when organisations collect fragments of information about GenAI use, they often cannot present those fragments as one coherent proof object. When a supervisor asks what happened in a specific run, when a complaint must be handled, or when an internal review is triggered, disconnected records are difficult to interpret and easy to contest. The evidence pack provides a structured answer.

It also avoids a common confusion between having data and having reviewable governance evidence. Logs, prompts, screenshots, approval emails, and version numbers may all exist, but if they are not assembled into a meaningful record, governance remains weak. The evidence pack converts scattered traces into an inspectable unit that can support explanation, challenge, and organisational learning.

If the evidence pack is missing, organisations may struggle to justify decisions, compare runs, explain why a score was assigned, or show that human oversight actually occurred. The result is often a return to assertion: people say a process was controlled, but cannot demonstrate it clearly. RAIDT uses the evidence pack to make operational governance visible.

Key idea: The evidence pack matters because it turns scattered run-level traces into a structured proof object that supports review, challenge, scoring, and governance readiness.

What this item captures

The identity of the run, including run identifier, task, actor or role, and timing.
The contextual basis of the run, including purpose, organisational setting, and any relevant constraints.
The generative process, including prompts, source materials, retrieval context, model choice, and settings.
The resulting artefacts, including outputs, revisions, and final approved versions where relevant.
Human oversight actions such as review, editing, escalation, approval, or rejection.
Evidence of integrity and control, such as checks performed, exceptions noted, and retention or redaction decisions.
The documented basis for producing or defending a RAIDT score profile.
The material needed for later reconstruction, contestability, and organisational learning.

Practical example / likely audience question

Audience question

Is the evidence pack just an archive of logs and documents, or does it do something more specific in RAIDT?

Answer

The concern behind this question is that organisations already generate many records, and a new governance artefact may look like extra bureaucracy. The direct answer is that the evidence pack is not simply a store of miscellaneous files. It is a structured, review-oriented package that assembles the evidential pieces of one run into a form that another person can inspect and use.

For example, imagine a financial-services team using GenAI to draft a customer complaint response. Raw traces may exist across several places: the prompt in one interface, the model version in system logs, the draft reply in a document, and approval comments in email or ticketing software. An evidence pack brings those pieces together and shows the run as one governance event. A reviewer can then see what the task was, what the model produced, how staff intervened, what checks were performed, and why the final response was accepted.

RAIDT handles this better than a generic AI governance approach because it does not stop at saying that evidence exists somewhere in the organisation. It requires a run-level proof object that can be reconstructed, inspected, and linked to scoring and governance readiness. That makes the evidence pack operational rather than merely archival.

Practical example in RAIDT terms

Consider a public-services setting in which a caseworker uses GenAI to draft a summary of a citizen's housing-support case before a supervisory review meeting. The GenAI use case is administratively useful, but the run-level issue is whether the summary accurately reflects the case record, avoids unsupported inferences, and can be defended if the citizen later challenges the decision process.

The evidence needed includes the run identifier, task purpose, prompt template, source case notes, any retrieval context, the model and configuration used, the generated summary, the caseworker's edits, the supervisor's comments, and the final decision about whether the draft could be used. Responsibility is affected because the organisation must show who reviewed and approved the summary. Auditability is affected because a later reviewer must be able to reconstruct the run. Interpretability is affected because the pack should show how the summary emerged from the prompt and source record. Dependability is affected because the organisation must assess whether the drafting process is consistently reliable. Traceability is affected because the run must remain linked to the relevant actor, artefacts, and decision stage.

The evidence pack improves governance readiness because it gives the organisation a defensible record for internal review, appeal handling, training improvement, and policy refinement. Instead of relying on a vague statement that staff used AI appropriately, the organisation can show what happened in one concrete case.

Detailed link to RAIDT

Evidence pack links to RAIDT in four ways.

First, it gives concrete form to RAIDT's core value of evidence over assertion by turning one GenAI use event into an inspectable governance artefact.
Second, it depends on the run and its run-level evidence, because the pack is assembled around one specific configured use of GenAI in context.
Third, it is one of RAIDT's two practical outputs and provides the documented basis from which a RAIDT score profile can be justified.
Fourth, it supports reviewability, contestability, audit readiness, and organisational learning because it gives reviewers a structured object for reconstruction and comparison.

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

Link to the five RAIDT pillars

Responsibility

The evidence pack supports Responsibility by showing who initiated the run, who reviewed it, who approved or rejected it, and what organisational purpose the run served.