simple_qwen3.5-397b_final2_la2
Qwen3.5-397B on the simple floor, 5 episodes. Cited by the thesis in:
| Wins (boss defeated, alive) | 4/5 [0.38, 0.96] |
| Final HP / exact optimum, mean over wins | 15% |
| Died | 0 |
| Final HP, mean | 111 |
| Seconds per model call | 91.2 |
| Parse error rate | 0.0% |
| Floor | Simple floor |
| Model | Qwen3.5-397B (qwen3.5-397b-a17b, thinking) |
| Planning | reactive |
| Prompt | engine-default-floor-basic |
| Temperature | 0 |
| Lookahead depth | 2 |
| Mechanical overrides | on (logged per turn) |
Engine settings (environment, 4)
THINKING_MAX_TOKENS=32768 CONVERSATION_HISTORY_TURNS=0 MOTA_ORDER_TWO_SIDED=1 MOTA_FIGHT_KEY_FLOOD=1
In a checkout of mt-ai-lab with an API key set (the result goes to a new file rerun_<time>_simple_qwen3.5-397b_final2_la2, never over the thesis result):
python scripts/run_catalog.py --codes G2
Result file: results/reliability/simple_qwen3.5-397b_final2_la2.json · updated 2026-09-27