edge width = summed absolute RELP flow between families
dashed = net signed sum is negative
cards show family label + highest-RelP member feature visualizations
hover a card to highlight groups connected through the aggregate RELP graph
Interpretation. This throws away individual-neuron clutter, but not the underlying accounting: each card says how many neurons were grouped, and each line is a sum over the detailed RELP edges cached from the 100 images. Toggle the icons between ordinary feature visualization and the fixed/path-guided feature visualization. Hovering a card propagates along the displayed aggregate RELP graph in both directions, so brighter cards are more strongly connected to the hovered group. The right column splits the old single sink into the most common predicted car-class logits in this cohort. The simplified view omits residual proxy/other buckets, while the detailed single-neuron view keeps those source neurons auditable.