Cell · cited by the thesis

complex_deepseek-v4-flash_survival_final2_plan5

DeepSeek-V4-Flash on the complex floor, 5 episodes. Cited by the thesis in:

Outcome
Wins (boss defeated, alive)4/5 [0.38, 0.96]
Final HP / exact optimum, mean over wins26%
Died0
Final HP, mean24
Seconds per model call43.7
Parse error rate0.0%
Settings
FloorComplex floor
ModelDeepSeek-V4-Flash (deepseek-v4-flash-0731, thinking)
Planningreactive
Promptcomplex_bfs_strategy_survival.txt
Temperature0
Lookahead depth3
Mechanical overridesoff (all engine replacements and assists disabled)
Engine settings (environment, 6)
MOTA_SUPPRESS_HINTS=0
THINKING_MAX_TOKENS=32768
CONVERSATION_HISTORY_TURNS=0
MOTA_ORDER_TWO_SIDED=1
MOTA_FIGHT_KEY_FLOOD=1
MOTA_PLAN_STEPS=5
Episodesopen one to replay it turn by turn
#outcomefinal HPHP / optstepsmodel callsreplay
1won2232%525replay
2won2232%515replay
3won2232%515replay
4not won49–559replay
5won69%525replay
Re-run this cell

In a checkout of mt-ai-lab with an API key set (the result goes to a new file rerun_<time>_complex_deepseek-v4-flash_survival_final2_plan5, never over the thesis result):

python scripts/run_catalog.py --codes DCP5

Result file: results/reliability/complex_deepseek-v4-flash_survival_final2_plan5.json · updated 2026-09-29