• workshop_dense_100#slide 61

Traceability in RAIDT is the ability to show, for one governed run, how an output connects to the materials, configurations, and decisions that produced it. The foundations paper defines it through provenance: what sources informed the output, what transformations occurred, and what system components contributed. RAIDT therefore does not treat traceability as a generic transparency slogan. It treats it as a concrete evidential property of a run-level evidence pack, recorded so that later reviewers can reconstruct the pathway from inputs and retrieved sources to output and oversight.

This matters because RAIDT is built around run as the unit of governance. In generative AI, apparently similar outputs may arise from different prompts, model versions, retrieval states, tool calls, or alignment settings. If those elements are not preserved, an organisation cannot reliably contest, audit, or learn from a specific case. Traceability also helps explain why RAIDT emphasises influence methods as governance interventions: these methods alter both behaviour and the evidence needed to review behaviour. When traceability is strong, the resulting score profile can distinguish a well-governed run from one that merely looks convincing. When traceability is weak, even an articulate output may be unusable for high-stakes reliance because the provenance chain cannot be checked. In this sense, traceability is central to contestability, audit support, compliance readiness, and organisational learning across repeated runs and changing configurations.

An organisation uses GenAI to help draft an employee performance appraisal. Later, the employee disputes the wording and asks what evidence and policy basis informed the draft. A traceable RAIDT run would let reviewers recover the prompt template version, any retrieved policy text, the model and deployment settings, the produced draft, and the manager's edits and approval decision before filing.

That concrete chain matters because the dispute is about one use event, not about the model in general. If the appraisal was drafted under a different prompt version, with different policy text, or after undocumented editing, the organisation needs to show that precisely. Without those records, it may still have a fluent output, but it does not have the provenance needed for fair review or defensible governance.

• 08-RAIDT_Foundations_M_V50
• 13-RAIDT-Evidence-Review_M_v10