Single-task execution with LLM-as-Judge verification in an iterative loop, supporting auto-retry and strict orchestrator role separation