Building Patchwork: What a 100-Defect Benchmark Taught Me About 3B Code Models
How an execution-guided LangGraph reflection loop boosted a 4-bit Qwen2.5-Coder-3B model from 38% Pass@1 to 58% Pass@overall, and the confirmation-bias blind spot it uncovered.