MT3-MT4_qwen3.5-397b_basic_d1d4
Qwen3.5-397B on the tower (MT3→MT4), 5 runs. Not cited by the thesis (development or earlier run).
| Highest floor reached, per run | MT4, MT4, MT4, MT4, MT4 |
| Died | 0 |
| Final HP, mean | 303 |
| Seconds per model call | 65.0 |
| Parse error rate | 1.6% |
| Floor | Tower MT3→MT4 |
| Model | Qwen3.5-397B (qwen3.5-397b-a17b, thinking) |
| Planning | reactive |
| Prompt | engine-default-tower-basic |
| Temperature | 0 |
| Lookahead depth | 3 |
| Mechanical overrides | on (logged per turn) |
Engine settings (environment, 1)
THINKING_MAX_TOKENS=32768
In a checkout of mt-ai-lab with an API key set (the result goes to a new file rerun_<time>_MT3-MT4_qwen3.5-397b_basic_d1d4, never over the thesis result):
python scripts/run_catalog.py --codes M4b
Result file: results/reliability/MT3-MT4_qwen3.5-397b_basic_d1d4.json · updated 2026-09-26