Cell

simple_deepseek-v4-flash_blockergraph

DeepSeek-V4-Flash on the simple floor, 5 episodes. Not cited by the thesis (development or earlier run).

Outcome
Wins (boss defeated, alive)1/5 [0.04, 0.62]
Final HP / exact optimum, mean over wins31%
Died0
Final HP, mean335
Seconds per model call12.1
Parse error rate0.0%
Settings
FloorSimple floor
ModelDeepSeek-V4-Flash (deepseek-v4-flash-0731, thinking)
Planningreactive
Promptblocker_lookahead.txt
Temperature0
Lookahead depth3
Mechanical overrideson (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_BLOCKER_GRAPH=1
LLM_TIMEOUT=1800
Episodesopen one to replay it turn by turn
#outcomefinal HPHP / optstepsmodel callsreplay
1not won249–2813replay
2won431%3617replay
3not won570–3616replay
4not won576–3617replay
5not won274–2813replay
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>_simple_deepseek-v4-flash_blockergraph, never over the thesis result):

python scripts/run_catalog.py --codes BGS

Result file: results/reliability/simple_deepseek-v4-flash_blockergraph.json · updated 2026-10-02