Tour · 6 of 11 · Sections 6.2–6.3 (RQ1)

The main result

Does the hybrid agent win where the heuristic and the bare model do not?

With the annotations, Qwen3.5-397B wins every one of its 15 episodes on the simple floor: ten earlier and five in the re-run on the final code, each at 2 HP [0.80, 1.00]. The rule-based heuristic, acting on the same engine output by fixed rules, ends the floor alive at 443 HP with the boss untouched. Without the annotations, the model loses too (0/3 per model, Table 6.3).

AgentSimpleComplexComplex, route in prompthand01hand02hand03hand04
exact optimum13 HP68 HP68 HP14 HP13 HP9 HP3 HP
Rule-based heuristic0/10/1—0/10/10/10/1
Qwen3.5-397B5/50/53/5 †15/54/50/50/5
DeepSeek-V4-Flash1/50/54/54/55/50/50/5
Qwen3.6-35B1/50/54/10 †34/53/50/50/5
Mistral Medium 3.50/50/50/5 †45/55/50/50/5

Wins k/N per cell of the re-run of Table 6.2; † = episodes in which the hero died. The heuristic row is quoted from Table 6.2.

Every model wins hand01 and hand02 in most episodes. No model wins the complex floor with the single-floor prompt, nor hand03 or hand04: their decisive choice is a long order of doors and fights. With the published solver's route written into the prompt, three of four models win the complex floor.

Hypothesis HA is supported for Qwen3.5-397B on the simple floor and for DeepSeek-V4-Flash on hand01 and hand02. Every cell links to its result file, its settings and its episodes.