• workshop_dense_100#slide 50

A tool chain trace is the part of the run-level evidence pack that records which tools were enabled for a run, which tool calls actually occurred, and the relevant outputs or returned artefacts that shaped the final model response. In the RAIDT papers, this sits chiefly within Traceability and Auditability: traceability anchors include tool chain IDs, while the technical foundation shows that a single output may depend on enabled tools, tool calls, retrieved passages, and external data sources. The trace therefore belongs to the evidence bundle for one configured use, not to abstract system documentation.

A concrete RAIDT-style example is a public-service eligibility assistant. A claimant query is processed under a specific prompt template, the model calls a search tool over the policy corpus, retrieves the current rule text, and drafts an explanation for a caseworker. The tool chain trace would record the enabled search capability, the relevant tool or retrieval identifiers, the snapshot IDs or hashes for the returned policy material, and the output that was reviewed or amended. Those fields let a reviewer see not only that a tool existed, but how it influenced this run.

This matters in RAIDT because governance claims must be inspectable at the point where risk materialises. Without a tool chain trace, a run may look plausible yet remain unreconstructable: reviewers cannot tell whether approved tools were used, whether external dependencies changed the answer, or whether the evidential basis of the output was stable enough to justify reliance. In scoring terms, absence or incompleteness of this trace shifts the run across the anchors 1=missing / 3=partial / 5=audit-ready and therefore changes the score profile across the five pillars (Responsibility, Auditability, Interpretability, Dependability, Traceability).

In a local-authority benefits team, a caseworker asks a GenAI assistant to draft advice on a residency rule. During that run, the system uses an approved search tool to retrieve the current policy clause and exemption guidance from the council policy store. A proper tool chain trace records that the search tool was enabled, the exact query or retrieval configuration, the snapshot identifiers for the returned passages, the passages themselves or their retained pointers, and the draft explanation shown to the caseworker.

If the resident later challenges the advice, reviewers can reconstruct whether the assistant relied on the correct policy version and whether the caseworker overrode or accepted the draft. That is why the trace matters in RAIDT: it turns a seemingly simple answer into a reviewable governance object rather than an untestable system memory.

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