simple_qwen3.5-397b_basic_multistep5_d1d4
Qwen3.5-397B on the simple floor, 3 episodes. Not cited by the thesis (development or earlier run).
| Wins (boss defeated, alive) | 0/3 [0.00, 0.56] |
| Final HP / exact optimum, mean over wins | – |
| Died | 3 |
| Final HP, mean | -6 |
| Seconds per model call | 106.1 |
| Parse error rate | 95.5% |
| Floor | Simple floor |
| Model | Qwen3.5-397B (qwen3.5-397b-a17b, thinking) |
| Planning | multistep-5 |
| Prompt | engine-default-floor-basic |
| Temperature | 0 |
| Lookahead depth | 3 |
| Mechanical overrides | on (logged per turn) |
Engine settings (environment, 1)
THINKING_MAX_TOKENS=32768
| # | outcome | final HP | HP / opt | steps | model calls | replay |
|---|---|---|---|---|---|---|
| 1 | died | -1 | – | 21 | 18 | log not stored |
| 2 | died | -15 | – | 17 | 14 | log not stored |
| 3 | died | -1 | – | 26 | 35 | log not stored |
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_basic_multistep5_d1d4, never over the thesis result):
python scripts/run_catalog.py --codes G2b
Result file: results/reliability/simple_qwen3.5-397b_basic_multistep5_d1d4.json · updated 2026-09-26