Prototype 01 · over-generate, cluster, then select

Behavioral clustering compresses breadth into a few representatives.

GenCluster abstraction · one dot = one candidate solution · the longest-thinking solution represents each behavior cluster · randomized pairwise comparisons rank clusters by representative wins · color shows ground-truth correctness, unseen during selection.

incorrect correct
Animated clustering of one hundred candidate solutions One hundred solution dots enter a dense grid, shake, and move into ten behavioral clusters. The longest-thinking solution represents each cluster. Randomized pairwise comparisons rank the clusters by representative wins, and all ten representatives fall into a submission row. Four are correct and six are incorrect.

Hundred Field L14 · Palm · GenCluster analogy 100 generated → 10 clusters → pairwise tournament → 10 submitted