Hidden overrides
Who really won the early episodes?
The early wins on the simple floor looked like the model's own. Then every stored log was checked turn by turn, comparing the model's answer with the action the engine executed.
In 94 turns of 68 stored episodes, the engine had executed a different action from the one the model chose, without logging it. A rule borrowed from the heuristic replaced the model's choice whenever it judged another action much cheaper. Every one of the 21 stored wins contained such a replacement; in 17 of them it came at turn 9, where the model chose the Slime Lord (1,9) and the engine opened the door (7,0) instead (Section 6.4.1).
- turn 9: the model chose Fight Slime Lord at (1,9) dmg=95, the engine executed Open yellow door at (7,0) episode lost
- turn 9: the model chose Fight Slime Lord at (1,9) dmg=95, the engine executed Open yellow door at (7,0) episode won
- turn 9: the model chose Fight Slime Lord at (1,9) dmg=95, the engine executed Open yellow door at (7,0) episode won
- turn 7: the model chose Fight Black Slime at (9,5) dmg=55, the engine executed Fight Bat at (3,5) dmg=182 episode lost
- turn 11: the model chose Fight Black Slime at (9,5) dmg=40, the engine executed Open yellow door at (4,7) episode lost
- turn 16: the model chose Fight Slime Lord at (1,9) dmg=75, the engine executed Open yellow door at (7,0) episode lost
Without the strategic override, the model lost every episode. Overrides were reduced to the mechanical cases and logged at every turn; in the fifteen winning episodes of Qwen3.5-397B in the final configuration, no override of any kind occurred. The lesson for evaluating such agents: log every intervention of the engine, or the wins may not be the model's.