S3.08 - Audit_trail

S3.08 ? Audit trail

flowchart LR
    A[Fragmented logs and disconnected records] --> B[RAIDT
Run-level evidence framework]
    A2[Weak reconstruction and limited challenge] --> B
    B --> C[[Audit trail
Linked evidential path for one governed run]]
    H[Healthcare, finance, education, cybersecurity] --> C
    I[Wrappers, metadata templates, review dashboards] --> C
    C --> D[Run-level evidence pack]
    C --> E[Five-pillar score profile]
    C --> F[Reviewer reconstruction and contestability]
    D --> G[Governance readiness]
    E --> G
    F --> G

? Star S3 - Run-Level Evidence Logic

Star context: Explains the proof-object logic of RAIDT by showing how a single run can be followed through prompts, configurations, tools, sources, outputs and review actions, so that evidence can support reconstruction, comparison and challenge.

Academic picture

Definition / background

An audit trail is the linked path through prompts, settings, tools, sources, outputs, review records, and subsequent decisions for a specific run. In ordinary information systems, the term often refers to logs that help show who did what and when. In RAIDT, the concept is narrower and stronger at the same time: narrower because it is anchored to one run as the unit of governance, and stronger because it is organised around evidential review rather than mere system activity.

This matters because generative AI use is frequently shaped by configuration choices, contextual instructions, source selection, human amendments, and approval steps that are not visible in a simple event log. A conventional platform log may record access time or API use, but still fail to explain why a particular output appeared, whether the run followed policy, or how a reviewer could challenge the result. RAIDT therefore treats audit trail as part of the proof-object logic of the framework: the trail helps connect evidence objects into a reviewable chain.

Conceptually, audit trail overlaps with provenance, traceability, lineage, and logging, but it is not identical to any of them. Provenance often focuses on origin; logging often focuses on events; lineage often focuses on data flow. RAIDT audit trail integrates these ideas around governance questions. It asks whether a reviewer can reconstruct a run, compare it with alternatives, assess control points, and see whether the evidence is sufficient for audit readiness.

Within RAIDT, the audit trail supports two practical outputs. First, it strengthens the run-level evidence pack by linking otherwise isolated records into a coherent evidential path. Second, it improves the credibility of the five-pillar score profile, because claims about responsibility, auditability, interpretability, dependability, and traceability are more defensible when the underlying trail can actually be inspected.

Why this concept matters

Audit trail solves a common governance problem in organisational GenAI use: many actors can say that controls exist, but few can show how a single use episode unfolded from instruction to output to review. Without that linked path, governance remains assertion-heavy. It becomes difficult to test whether the run followed policy, whether a harmful result could be contested, or whether lessons from one case can improve later practice.

The concept also avoids a major confusion. People often assume that retaining platform logs, chat history, or version records is enough. In practice, those records may be incomplete, scattered, or disconnected from governance review. RAIDT makes audit trail useful by tying it to the questions an organisation must answer about one run: what was attempted, under which conditions, with which evidence, who reviewed it, and what governance implications followed.

If audit trail is missing, important risks appear quickly. Reviewers may be unable to reconstruct a contested decision. Teams may fail to distinguish a prompt problem from a source problem or a reviewer problem. Score profiles may look neat on paper but rest on weak evidence underneath. In that sense, audit trail is one of the mechanisms that moves GenAI governance from principles to operational scrutiny.

Key idea: Audit trail matters because RAIDT turns fragmented technical traces into a run-level evidential path that supports reconstruction, challenge, scoring, and audit readiness.

What this item captures

The sequence of actions that constitute one governed run.
The relationship between prompts, settings, tools, sources, outputs, and human interventions.
The checkpoints at which review, approval, escalation, or correction occurred.
The evidential basis for reconstructing why an output was produced in a particular context.
The material needed to justify a run-level evidence pack and defend pillar scores.
The trail needed for contestability, organisational learning, and policy refinement.

Practical example / likely audience question

Audience question

How is it different from generic logs?

Answer

The concern behind this question is that organisations already collect logs, timestamps, and activity records, so audit trail may sound like old terminology for existing infrastructure. The direct answer is that RAIDT audit trail is not just a store of events. It is a structured governance pathway organised around one run and around the questions that reviewers, supervisors, auditors, and policy actors need to answer.

A generic log might show that a user accessed a tool at 10:14, called a model, and generated an output. RAIDT audit trail goes further. It links the task purpose, prompt wording, model or tool settings, source materials, intermediate transformations, output version, reviewer comments, approval status, and the implications for the five-pillar score profile. That makes the trail useful for reconstruction and challenge rather than mere operational monitoring.

This is where RAIDT improves on generic AI governance approaches. Many governance schemes state that organisations should retain records, but they do not specify how those records become a reviewable run-level evidence chain. RAIDT supplies that operational structure. It helps a reviewer examine not only whether records exist, but whether the records answer governance questions in a coherent and contestable way.

Practical example in RAIDT terms

Consider a healthcare trust using a generative AI assistant to draft a discharge summary from clinician notes. The run-level issue is not simply that the model produced text. The issue is whether the organisation can later show what notes were supplied, which prompt template was used, whether medication instructions were checked, whether a clinician edited the draft, and whether the final version matched clinical policy.

The evidence needed includes the prompt template, model version or system configuration, input-note references, generated draft, clinician edits, approval record, and any escalation note if the output was judged unsafe or incomplete. In RAIDT terms, this evidence forms the audit trail for that run.

The most affected pillars are Auditability and Traceability, but Responsibility and Dependability are also involved because clinical review duties and output reliability depend on the quality of the trail. By improving the audit trail, the organisation becomes more governance-ready: it can reconstruct a contested discharge summary, identify where a failure occurred, and demonstrate that human oversight and evidential checks were not merely assumed but documented.

Detailed link to RAIDT

Audit trail links to RAIDT in four ways.

First, it operationalises RAIDT's core idea that governance should attach to a specific run rather than to broad claims about a system in general.
Second, it makes run-level evidence inspectable by connecting the component records of that run into a coherent path.
Third, it strengthens both the evidence pack and the score profile, because each depends on evidence that can be followed, interpreted, and checked.
Fourth, it supports reviewability, contestability, audit readiness, and organisational learning by showing how a run unfolded and where intervention was possible.

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

In that chain, audit trail is the connective mechanism. It does not replace evidence objects, but it makes them usable as a proof structure rather than a loose archive.

Link to the five RAIDT pillars

Responsibility

Audit trail clarifies where responsibility sat during a run, including who initiated the task, who reviewed the output, and who authorised use or correction. It helps separate accountable human decisions from automated generation steps.