{"agent_comparison":{"description":"Mean scores across 10 tasks, 20 seeds each","agents":{"random":{"avg":0.054,"description":"Random action selection — lower bound"},"rule_based":{"avg":0.092,"description":"Scripted heuristic agent"},"information_first":{"avg":0.536,"description":"Evidence-gathering heuristic — reads before acting"},"grpo_llama":{"avg":0.53,"description":"GRPO fine-tuned Llama (see train_grpo.ipynb)"}}},"benchmark_table":{"easy":{"random":0.138,"rule_based":0.713,"info_first":0.877},"medium":{"random":0.083,"rule_based":0.001,"info_first":0.546},"hard":{"random":0.062,"rule_based":0.2,"info_first":0.15},"bonus":{"random":0.065,"rule_based":0.001,"info_first":0.999},"security":{"random":0.006,"rule_based":0.001,"info_first":0.05},"database":{"random":0.042,"rule_based":0.001,"info_first":0.15},"failover":{"random":0.002,"rule_based":0.001,"info_first":0.001},"dns":{"random":0.036,"rule_based":0.001,"info_first":0.78},"waf":{"random":0.09,"rule_based":0.001,"info_first":0.807},"thundering_herd":{"random":0.016,"rule_based":0.001,"info_first":0.999}},"note":"Run python baselines.py to reproduce. GRPO scores from train_grpo.ipynb."}