evolution-looplisted

# Evolution Loop ## When This Skill Applies - Understanding how Artibot learns and self-improves over time - Configuring the learning pipeline (nightly schedule, thresholds, GRPO settings) - Interpreting GRPO training results or pattern extraction outputs - Planning how to extend or customize the evolution loop for your workflow ## Core Guidance ### What Is the Evolution Loop? The evolution loop is Artibot's autonomous improvement cycle. It continuously extracts patterns from usage, ranks them by effectiveness, trains the model's preferences (GRPO), and promotes the best patterns into System 1 (fast, intuitive responses). ``` Session Data → Pattern Extract → Quality Score → GRPO Training → Knowledge Update | System 1 Promotion ← Skill Refinement ← Swarm Merge ←----------------+ ``` ### Five Stages | Stage | Description | Output | |-------|-------------|--------| | **1. Self-Scan** | Analyze recent session data: tool usage, errors, team compositions | Raw pattern candidates | | **2. Pattern Extract** | Score candidates by frequency, success rate, and novelty | Ranked pattern list | | **3. Knowledge Update** | Merge high-confidence patterns into the knowledge base | Updated knowledge store | | **4. Skill Refinement** | Auto-update SKILL.md files with improved guidance based on real usage | Refined skill content | | **5. GRPO** | Group Relative Policy Optimization — train preference ranking across resp