In a contract dispute, an agent may use a current amendment, exclude a superseded clause, and encounter an uncertain transcript. Traceability makes the resulting position reconstructable and defensible: reviewers can inspect the evidence and decisions that shaped it, including how a supervising lawyer changed or approved the draft. section 5.6
Traditional lineage follows data origins and movement. Agentic traceability also records how a decision episode was assembled and what happened afterward. A generated explanation is not evidence; an independently inspectable record must support it. section 5.6
Connect the evidence for each decision episode:
- Record sources retrieved and excluded, transformations, graph traversals, policy decisions, model and configuration versions, tool calls, outputs, reviews, and final disposition. section 5.6 section 5.8.6
- Propagate a common decision episode identifier so records can be correlated without relying on the model’s explanation. Retain the identifiers and versions needed for reconstruction. section 5.8.6
- Trace each material policy obligation to its source, owner, applicability conditions, evaluation result, exception route, and runtime response. Enforce it through guardrails or workflow controls rather than prompt instructions alone. section 5.7.4
Reconstruct and respond to change
Reconstruct a preliminary position on breach and damages from independent records: the governing agreement, evidence used and excluded, authorities, policy decisions, agent actions, and lawyer review. Then change a material condition during an active episode. Confirm that the solution identifies affected context and derivatives, prevents continued reliance, re-evaluates authority, and records its response. section 5.8.6
Operational traceability also depends on continuity. A source correction, permission change, policy update, or other invalidation trigger should initiate an approved response, such as refresh, restriction, rebuild, authority reduction, suspension, or transfer. section 5.8.6



