complex_deepseek-v4-flash_exact68_final2
DeepSeek-V4-Flash on the complex floor, 5 episodes. Cited by the thesis in:
| Wins (boss defeated, alive) | 1/5 [0.04, 0.62] |
| Final HP / exact optimum, mean over wins | 59% |
| Died | 0 |
| Final HP, mean | 55 |
| Seconds per model call | 9.6 |
| Parse error rate | 0.0% |
| Floor | Complex floor |
| Model | DeepSeek-V4-Flash (deepseek-v4-flash-0731, thinking) |
| Planning | reactive |
| Prompt | complex_exact68_sequence.txt |
| Temperature | 0 |
| Lookahead depth | 3 |
| Mechanical overrides | off (all engine replacements and assists disabled) |
Engine settings (environment, 6)
MOTA_SUPPRESS_HINTS=0 THINKING_MAX_TOKENS=16384 CONVERSATION_HISTORY_TURNS=0 MOTA_ORDER_TWO_SIDED=1 MOTA_FIGHT_KEY_FLOOD=1 LLM_TIMEOUT=300
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_exact68_final2, never over the thesis result):
python scripts/run_catalog.py --codes X68D X68D1
Result file: results/reliability/complex_deepseek-v4-flash_exact68_final2.json · updated 2026-09-29