Evaluation
Trajectory evaluation
Evaluation of the full agent run, including tool calls, risk classes, approval events, and outcomes.
Trajectory evaluation assesses the path an agent took, not only the final answer. It reviews tool sequence, state, risk classification, approval boundaries, sensor results, and final output.
Why it matters
A final answer can look correct even if the agent skipped a required control. Trajectory evaluation catches governance failures.
How HARNEXA uses it
HARNEXA uses trajectory evidence in ORACLE, CLEAR scorecards, PHANTOM proof, and AgentOps reporting.
FAQ
Questions this definition answers.
Why not evaluate only the final output?
Because governance failures often happen in the middle of the run: wrong tool, wrong risk class, missing approval, or missing audit event.
What does trajectory evidence include?
Tool calls, timestamps, state transitions, risk classes, approval events, sensor results, and run outcomes.
Need these definitions mapped to a live commerce workflow, agent estate, or protocol-readiness path?
Book diagnostic sprint