Iterative self-improvement system with task complexity grading, strict quality gatekeeper role, confidence thresholds, and verification checklists

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