In every figure you have met so far, the coupling K was a knob you slid — one number, set from outside, the same for every pair. This section takes the knob away. The coupling becomes a matrix: a number Kij for every pair of oscillators, and nothing stops that matrix from being different for every pair and different at every instant. The most interesting choice — the one this whole site keeps circling — is to let the states themselves decide it. When agreement chooses the wiring and the wiring shapes the agreement, the network is not running on a graph. It is continuously growing one.
Write down who pulls on whom and how hard, and you get a square grid of numbers: row i lists how strongly oscillator i listens to each of the others. A fixed, uniform K is the special case where every entry is the same. A two-community graph is the case where the grid has two blocks and silence elsewhere. And an effective coupling is the general case: the entries are computed, moment by moment, from the current states — for instance, "listen most to the five you currently agree with most."
That last rule creates a loop with consequences. Agreeing makes you couple; coupling makes you agree harder. Rich-get-richer, in geometry. Below, watch both halves of the loop at once: the sphere on the left, and the actual K matrix on the right, drawn live. Early on, the wiring flickers — everyone's nearest agreements keep changing as the crowd churns. As clusters condense, the wiring freezes into structure: each oscillator's chosen five stabilize into its clustermates. The matrix stops flickering because the geometry stopped churning — and the geometry stopped churning because the matrix chose it. Neither came first. That is what "the wiring chooses itself" means.
Here is the experiment that shows why any fixed picture of the coupling is an approximation. Two identical populations run side by side with agreement-chosen wiring. Press Quench the right sphere: from that instant, the right population keeps the wiring it happens to have — the same partners forever, a snapshot of a living thing — while the left keeps re-choosing every moment. They look identical. They are not, and one shove reveals it.
Press Kick both: the same random shove scatters both populations. For a few seconds, look at the bonds — the lines showing who is wired to whom. The living network's bonds stay short: the instant its members landed, each re-chose the five nearest agreements, so its wiring always matches its geometry. The frozen network's bonds are stretched across the sphere: each member still pulls toward partners from the vanished arrangement, wherever the kick threw them — and it spends those seconds hauling its old groups back together across the surface.
Then wait, and read the meters. Both networks settle into tidy clusters again — but not the same clusters. The meter under each sphere scores how much of the pre-kick arrangement survived: which pairs that were clustermates before are clustermates again. The frozen network scores high — it has rebuilt its past, because its wiring is its past, written down. The living network scores low — it organized whatever the new scatter suggested, with no loyalty to the old grouping. That is the whole distinction in one experiment: a frozen coupling is a memory; a living coupling is a sense organ.
This site's research discipline has a blunt rule, stated in its working notes and worth quoting in spirit: the coupling field is dynamical; every fixed-K computation is a quenched approximation and must be labeled as one. You have now felt why. A snapshot of a living network's wiring is not the network — it is where the network happened to be standing. Any analysis done on the snapshot inherits that staleness the moment the states move.
And you have seen this loop before, wearing its most famous clothes: in the transformer primer, attention computes a fresh who-listens-to-whom matrix from Q·K agreements — per token, per layer. On this site's reading, attention is an agreement-chosen effective coupling recomputed at every step: the wiring choosing itself, industrialized. The Transformers essay develops and measures that identification on real models; the next section follows a single token through exactly this kind of living wiring.