Where do attention and MLP updates go after each sublayer?

A shared page gathers useful notes

Imagine every token position carrying a working page through the transformer. The page begins with that token’s initial vector. Each attention layer can add a note gathered from other positions, and each MLP can add a locally computed note. The page is the shared channel through which components communicate; it is not readable prose. Its size stays the same while its contents become increasingly contextual. For next-token prediction, the final page at the last prompt position goes to the language-model head. We can project earlier pages into vocabulary space to watch candidates rise and fall, but that projection is an interpretability probe. It helps form hypotheses about how a prediction develops; it is not a recording of private thoughts.

Every block adds updates to temporary token states that keep the same shape; those updates are not automatically evidence in the factual sense.