• qa_deck_100#slide 44 · Retrieval, tools, adapters, and alignment evidence

In RAIDT, a tool chain trace counts as governance evidence because it bears directly on whether a disputed run can be reconstructed, reviewed, and contested. The papers argue that GenAI risk materialises at run time: outputs are shaped not only by the model, but by prompts, retrieved context, enabled tools, tool calls, external data sources, and oversight actions. For that reason, RAIDT treats the run as the unit of governance and requires a run-level evidence pack rather than relying on system descriptions or generic telemetry. A trace of tools used in one run is therefore not merely descriptive metadata about software operation. It is evidence about how a materially consequential output was produced under specific controls, with specific dependencies, at a specific time.

The distinction from engineering metadata is central to the technical foundation. Engineering logs are valuable, but the papers repeatedly note that raw traces are often fragmented, technically shaped, and unintelligible to organisational reviewers. They may support observability, yet fail to support managerial, compliance, audit, procurement, or dispute-resolution review unless they are assembled into a bounded governance object. In RAIDT, the tool chain trace becomes governance evidence when it is preserved with identifiers, hashes, versions, and oversight context so that it can inform the five pillars (Responsibility, Auditability, Interpretability, Dependability, Traceability), contribute to the score profile, and be judged against the anchors 1=missing / 3=partial / 5=audit-ready. Because tool access is one of the influence methods as governance interventions, the trace is part of answerability, not incidental telemetry.

A local authority uses a GenAI assistant to draft eligibility advice for a housing-benefit case. The assistant queries a search tool over the current policy corpus and returns a suggested explanation to a caseworker. If the claimant later challenges the advice, an engineering dashboard showing service uptime or aggregate search latency is not enough. Governance review needs the tool chain trace for that run: which search capability was enabled, which policy snapshot was queried, which clauses were returned, whether the model used those clauses, and which officer approved the final communication.

That record belongs in the run-level evidence pack because it supports reconstruction and contestability. It lets reviewers test whether the system operated under approved controls, whether the advice relied on the correct policy version, and whether human oversight was exercised. Without it, the organisation has metadata about infrastructure, but not governance evidence about the decision-support event.

• 08-RAIDT_Foundations_M_V50
• 13-RAIDT-Evidence-Review_M_v10
• 18-RAIDT-Technical-Foundation_M_v04