We knew what we wanted to preserve before we knew how to draw it: take the MAE parcel associated with a neuron, send its members across layers, and keep the individual routes legible. The interface took a few attempts to understand the assignment.
The first map technically contained the relationships. This was not the same as being able to read them. Hundreds of routes were present; almost none of their structure survived contact with the eye.
Useful failure: the data was there. The representation was not.
We forced the routes into layers and columns. Suddenly there was an origin, destinations, and something resembling a map. It was less beautiful than the mess and considerably more useful.
L5:N541 begins to look less like a starburst and more like a source feeding traces.
The map learned to breathe. Layer headings, aligned nodes, a persistent detail pane. You could finally follow individual paths without mentally reconstructing the whole system from spaghetti.
The temptation now was to simplify harder. This turned out to be dangerous.
At some point the interface became tidier by collapsing precisely the distinctions we cared about. So I drew the requirement again: one source, one parcel, many members, many destinations, individual routes preserved.
Approximately one evolutionary step above cave painting. Extremely effective.
Once the requirement was explicit, the geometry changed. Straight runs. Junctions. Branches. Components arranged by layer. Shared structure where routes agree; divergence where they do not.
This was the point where the visual language began solving the problem for us.
The thread map returned to the game as a compact instrument. The routes were no longer trying to resemble physical space. They were traces: a readable schematic of relationships across components.
We had spent the morning designing circuit diagrams without noticing.