We investigate where a Transformer shifts from memorization to generalization. To probe this, we introduce Transition Games, exact activation games aligned with model behavior and paired with non‑generalizing controls. Adding a prospective bias to block‑0 attention yields a distributed utility gain; degree‑two modes alone account for 67%–92% of the added contrast in replacement experiments. An exact path study shows that the block‑1 MLP mediates a larger share of this effect than any downstream path across all 12 tested pairs.
The conventional claim “MLP memorizes, attention generalizes” is reversed in our findings: information per example drops by $.331$ bits at the memory anchor, and all 12 pairs move opposite to the predicted direction. Hypotheses such as routing onset, global rank collapse, and a prime‑invariant architectural ridge also fail to explain the results. Together, the evidence points to grokking manifesting as a spectral (Fourier) recoding of an existing distributed circuit rather than a switch between modules.
Review