Q, K and V start identical, seeded with random integers 1–6. Raw scores (Q × Kᵀ), attention weights (row-wise softmax of the scaled scores) and context vectors (Z) are computed from those values. Moving a slider re-seeds Q, K and V.