Why this shape, not a helix. A double helix is two distinct strands bridged by exclusive one-to-one rungs. Attention is one sentence crossed with itself — the same tokens on both axes, connected many-to-one with graded weights. The honest topology is a complete weighted directed graph: every token can attend to every earlier token, each edge carrying a continuous weight. The 2D panel flattens that graph onto a ring — chords woven through the interior, colored by kind. The orange fan is the sink (every query reaches token 0); the green short chords are recency (neighbors attending to neighbors); grey is everything else.
Your inverse-helix, made literal. The 3D view takes your image at its word: turn the helix inside-out so the strands run through the interior of the circle. Query tokens ride the top ring, key tokens the bottom ring (the same sentence twice, in its two roles — looker and looked-at), and each attention weight is a strand bowing through the central axis from top to bottom. The sink becomes unmistakable: every query's strongest strand dives to the same point on the key ring — a convergent bundle, the funnel your eye kept seeing, now honestly a many-to-one gather rather than a pipe. A slight twist between the rings gives it the helical wind you pictured. Drag to rotate; toggle HIGHLIGHT to isolate one query's strands, ALL to see the full mesh.
Amber tier: the weights are modeled to reproduce documented attention structure (sinks, recency, heavy tail), not extracted from a trained network — but the topology is exact. This is genuinely how a causal attention layer connects: a self-crossed weighted graph, not two strands, not a funnel.