simple_deepseek-v4-flash_selflookahead
DeepSeek-V4-Flash on the simple floor, 3 episodes. Not cited by the thesis (development or earlier run).
| Wins (boss defeated, alive) | 0/3 [0.00, 0.56] |
| Final HP / exact optimum, mean over wins | – |
| Died | 0 |
| Final HP, mean | 343 |
| Seconds per model call | 84.4 |
| Parse error rate | 9.8% |
| Floor | Simple floor |
| Model | DeepSeek-V4-Flash (deepseek-v4-flash-0731, thinking) |
| Planning | reactive |
| Prompt | self_lookahead.txt |
| Temperature | 0 |
| Lookahead depth | 3 |
| Mechanical overrides | on (logged per turn) |
Engine settings (environment, 7)
THINKING_MAX_TOKENS=32768 CONVERSATION_HISTORY_TURNS=0 MOTA_ORDER_TWO_SIDED=1 MOTA_FIGHT_KEY_FLOOD=1 MOTA_BARE_OBS=1 MOTA_GRID_MAP=1 LLM_TIMEOUT=1800
In a checkout of mt-ai-lab with an API key set (the result goes to a new file rerun_<time>_simple_deepseek-v4-flash_selflookahead, never over the thesis result):
python scripts/run_catalog.py --codes SLS
Result file: results/reliability/_superseded_grid/simple_deepseek-v4-flash_selflookahead.json · updated 2026-10-01