Cell

tower_deepseek-v4-flash_final2_maxsteps

DeepSeek-V4-Flash on the tower (MT0→MT23w), 3 runs. Not cited by the thesis (development or earlier run).

Outcome
Highest floor reached, per runMT3, MT4, MT3
Died0
Final HP, mean65
Seconds per model call9.8
Parse error rate0.0%
Settings
FloorTower MT0→MT23w
ModelDeepSeek-V4-Flash (deepseek-v4-flash-0731, thinking)
Planningreactive
Promptengine-default-tower-basic
Temperature0
Lookahead depth2
Mechanical overrideson (logged per turn)
Engine settings (environment, 6)
THINKING_MAX_TOKENS=16384
CONVERSATION_HISTORY_TURNS=0
MOTA_ORDER_TWO_SIDED=1
MOTA_FIGHT_KEY_FLOOD=1
MOTA_PLAN_STEPS=60
LLM_TIMEOUT=1800
Runsopen one to replay it turn by turn
#outcomefinal HPHP / optturnsmodel callsreplay
1MT325–2610replay
2MT4146–6635replay
3MT325–3212replay
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>_tower_deepseek-v4-flash_final2_maxsteps, never over the thesis result):

python scripts/run_catalog.py --codes DTM DTMa

Result file: results/reliability/tower_deepseek-v4-flash_final2_maxsteps.json · updated 2026-10-01