One block pair that has a flagged superparent: B0→B1, B1→B2 — 2 of the 3 pairs graded in this run. The left node is the parent feature; the right nodes are up to 25 of its children, ordered by the strength of the link.
Pairs with no diagram
B2→B3 is graded but no parent there reaches the flag, so it gets no diagram. The gate is fan-out alone, at SUPERPARENT_OUTDEG_FRAC = 30% of the child block. B3→B4 is a different case: not graded at all, so absent rather than empty. It is the largest co-firing matrix in the run and is off unless EXP0_B3B4=1, which needs stage 01 re-run, not just a redraw.
Ribbon width
The reverse coverage of that one parent→child edge — P(parent fires | child fires), the quantity the edge criterion (≥ 0.5) is applied to. Width is per edge; it is not a share of the parent.
Ribbon colour
The verdict on that edge. Where stage 03 has run the colour is the probe-based S_res rank test (both decoders inside the top k = 5 of the child probe's correlations); where it has not, it falls back to the reconstruction and frequency filters. The legend above the first diagram names whichever is in force.
Node labels
The feature index, plus its Neuronpedia label when feature_labels.json is present. Bare indices mean the labels were not fetched for this run, not that the features are unlabelled.