Training worked
3 / 3 seeds
Validation improved, parameters moved, and every checkpoint stayed within 2.6% of Lloyd-10.
2jw5U060C4 · three trained seeds · paired held-out and OOD controls
Validation improved, parameters moved, and every checkpoint stayed within 2.6% of Lloyd-10.
Normalized costs: trained 0.64676, Lloyd-10 0.64306; paired CI crosses zero.
Trained 1.03585 vs Lloyd-10 1.01574; 95% CI [0.01119, 0.03040].
The paper uses chosen parameters without training and expressly leaves learnability to future work.
The challenge claim is falsified in this executed protocol: successful optimization does not produce superior clustering or OOD generalization. This does not prove impossibility for every conceivable learning environment.