flow

vertical residual stream · 30 blocks top-to-bottom · each block tees off twice (attention add, ffn add) · sparklines are truncated singular value spectra of the block's composed maps, plotted on a shared log-σ axis · hover any block for stats

small multiples

same data as the flow above, organized for layer-to-layer comparison · all sparklines share one global log-σ y-axis · the unevenness in column 1 is the SWA/global periodicity