AI Operations · Glossary
Per-Decision Traceability
Architectural requirement 1 of an Interactive Agent. Per-Decision Traceability means every decision the agent makes is inspectable down to the Policy, the Routine step, and the context that produced it.
Per-Decision Traceability is the substrate of three operational behaviours that an Interactive Agent must support. First, audit response: when a regulator, an internal compliance team, or a customer challenges a decision, the operator can produce the full record of why the decision happened. Second, evaluator input: EvalOps runs evaluations against decisions, not against transcripts; the trace is the evaluator's input. Third, the autonomous improvement loop: Co-Pilot reads production traces to surface patterns and draft refinements. Without per-decision granularity, none of the three behaviours hold. Session-level logs or model-call logs are not sufficient. The unit of traceability is the decision, recorded in the production trace.