Tour · 9 of 11 · Sections 6.3, 6.4

Where episodes are lost

At which turn does a losing episode leave the winning one, and how far ahead must the engine look?

Comparing each failed episode with the winning trajectory action by action locates the first departure, and the state value after it shows whether that departure lost the floor. In the re-run of Table 6.2, ten of the thirteen failures depart at one of two turns.

At turn 18, with 720 HP, ATK 10 and DEF 6, only the Bat (3,5) at 132 HP keeps the boss winnable; the Slime Lord (4,2) costs 60 HP now against 22 later. At turn 22, the door (7,5) keeps the value at 2 HP; the Slime Lord (4,2) costs 52 now against 22 later, and the boss can no longer be defeated.

Qwen3.5-397B without conversation history: a win and a loss of the same cellIdentical for 17 turns. At turn 18 the win takes Fight Bat at (3,5) dmg=132, the loss Fight Slime Lord at (4,2) dmg=60.State value before: 2 HP; after: win 2 HP, loss lost.compare the two episodes

Qwen3.6-35B, final configuration: a win and a loss of the same cellIdentical for 21 turns. At turn 22 the win takes Open yellow door at (7,5), the loss Fight Slime Lord at (4,2) dmg=52.State value before: 2 HP; after: win 2 HP, loss lost.compare the two episodes

How far ahead does the engine have to look for the model? The lookahead tree shows what an action unlocks up to three gateways deep. Cutting it shows that the depth is load-bearing on this floor:

Qwen3.5-397BQwen3.6-35B

depth 3 (final)10/107/10

depth 24/57/10

depth 10/30/5

depth 00/30/5

Lookahead depth on the simple floor (Table 6.3). At depth 1, all eight episodes open the yellow door (7,5) early; at depth 0, five of eight fight the Slime Lord (1,9) before the stat item that makes it cheaper.

The depth is floor-specific: on hand01 and hand02, DeepSeek-V4-Flash wins 9 of 10 episodes at depth 1, as at depth 3.