Defects are first-class entities, not notes buried in a test record. When the agent hits a failure during execution, the defect is logged immediately with the step, the expected versus observed result, and the evidence. Each defect moves through a structured six-state lifecycle, new, triaged, in-progress, resolved, verified, and closed, and carries a root-cause classification across five branches. Resolving a defect triggers re-testing of the affected step, so a fix is confirmed rather than assumed.