Induction head. The defining mechanism of in-context learning. On a repeated sequence, at each token it searches back for the previous occurrence of that same token and attends to whatever followed it — then copies that as the prediction. The heatmap shows the resulting off-diagonal induction stripe; the copy-score measures how often the copied token is actually the correct next token. Amber tier: the match-then-copy algorithm is exact, the sequence is constructed to isolate it (real induction heads need a model large enough to have formed them in training).