SGK86/procedural-graphs-repro

fix: missing import time in run_batch

report: reproduction of arXiv 2609.09153 on HotpotQA with Qwen2.5-72B-Instruct

main: baseline eval -> skeleton val -> 3 evolution rounds -> final eval

HotpotQA EM/F1 metrics + pool loading

offline self-evolution: diagnose -> refine -> validation gate -> rejection memory

LLM backend via HF Inference with model fallback + retry

guided agent: locate -> 2-hop extract -> generate guidance -> solve

procedural graph structure + edit operations

initial commit