{"label":"simple_qwen3.5-397b_basic_multistep5_d1d4","config":{"floor":"simple","start":null,"end":null,"max_turns":null,"checkpoint":null,"hero_override":null,"prompt_file":"engine-default-floor-basic","prompt_chars":2527,"model":"qwen3.5-397b-a17b","task_mode":"thinking","preset":"qwen3.5-397b","mode":"multistep-5","temperature":0,"lookahead":3,"structured_output":true,"overrides_disabled":false,"env":{"THINKING_MAX_TOKENS":"32768"},"source":"cli"},"runs":[{"score":0.6172,"goal_score":0.0964,"success":false,"hp_ratio":null,"boss_fight_won":true,"goal":"boss","goal_reached":true,"died":true,"final_hp":-1,"hero_atk":11,"hero_def":6,"enemies_killed":9,"enemies_total":14,"steps":21,"violations":["suicide_fight: Boss at (5,10)"],"llm_calls":18,"parse_errors":18,"forced_turns":0,"run_log":null,"elapsed_s":1201,"timestamp":"2026-09-26T04:14:37","run_idx":1},{"score":0.2673,"goal_score":0.075,"success":false,"hp_ratio":null,"boss_fight_won":false,"goal":"boss","goal_reached":false,"died":true,"final_hp":-15,"hero_atk":11,"hero_def":5,"enemies_killed":7,"enemies_total":14,"steps":17,"violations":["suicide_fight: Slime Lord at (1,9)","boss_not_killed"],"llm_calls":14,"parse_errors":14,"forced_turns":0,"run_log":null,"elapsed_s":2465.6,"timestamp":"2026-09-26T04:55:43","run_idx":2},{"score":0.6464,"goal_score":0.1071,"success":false,"hp_ratio":null,"boss_fight_won":true,"goal":"boss","goal_reached":true,"died":true,"final_hp":-1,"hero_atk":11,"hero_def":6,"enemies_killed":10,"enemies_total":14,"steps":26,"violations":["suicide_fight: Boss at (5,10)"],"llm_calls":35,"parse_errors":32,"forced_turns":0,"run_log":null,"elapsed_s":3441.1,"timestamp":"2026-09-26T05:53:04","run_idx":3}],"aggregate":{"n":3,"successes":0,"success_rate":0,"success_ci95":[0,0.5615],"died":3,"hp_ratio_mean":null,"hp_ratio_sd":null,"goal_score_mean":0.0928,"goal_score_sd":0.0163,"score_mean":0.5103,"score_sd":0.211,"score_min":0.2673,"score_max":0.6464,"final_hp_mean":-5.6667,"final_hp_sd":8.0829,"llm_calls_total":67,"seconds_per_call":106.09,"parse_error_rate":0.9552,"elapsed_s_total":7107.7},"updated":"2026-09-26 05:53"}