A good method
knows its limits.
One rule. Many landscapes.
See where a generated method helps, where it loses, and what the evidence actually supports.
Loading the recorded experiments…
Two frozen methods. No automatic promotion.
Campaign 1 attempted 256 proposals. Campaign 2 attempted 4,096, including duplicates. Neither selected method met all of its declared acceptance requirements. The reference uses nearest-neighbor construction followed by fixed 2-opt refinement; it is a heuristic, not an optimum guarantee.
A counterexample you can inspect.
The recorded cases contain the supplied points and both closed tours. The interactive replay shows those existing routes. It performs no new scientific calculation.
Different geometries, different outcomes.
Compare clusters, corridors, rings and other synthetic families. A favorable mean does not rule out individual losses. Published results describe the completed campaigns, not a newly validated method-selection rule.
A correction needs new evidence.
A counterexample used to improve a method becomes development data. Freeze the next proposal and its acceptance rules before evaluating it on fresh cases.
Read the complete recorded evidence