The TSP role-split made visible — optimal solver route vs. direct model route.
Each tab shows one Arizona fixture — fixed Euclidean points with real Arizona city names. The green solid route is what the deterministic solver produces: always optimal. The red/amber route is what a direct model generates.
The greedy trap (tsp-003) is the most instructive: the direct route looks plausible but quietly drops Tucson, making it not a valid tour. The grayed city marked ✗ is the one never visited.
The role-split is clearest in tsp-003. The direct model closed a tour without visiting Tucson — structurally invalid. The orchestrated path never fails this way because correctness belongs to the solver, not the model.
On tsp-001 and tsp-004, a strong model can match the optimal route exactly. That is the best-case direct result. It does not generalize to harder instances, and weaker models fail structurally on tsp-004.