In the memory chapter, the carved basin was drawn as a pull toward a stored pattern — a deliberate cartoon, and that chapter says so: the real basin, its prose promises, lives in the slow part of the coupling field. This section keeps that promise. Here the memory is stored nowhere except in K — strong bonds where the pattern agrees with itself, dead bonds where it doesn't. Which means the memory can be damaged the way real substrates are damaged: bond by bond. And it means something a stored array can never do: with the learning rule running, the memory can grow back.
The lattice below stores a two-tone flag: left half one phase, right half the opposite. The storage is pure K-field — you can see it as the web of bright bonds inside each half. The cycle to run: Probe measures recall (give the network a 45%-corrupted cue, let it settle with the wiring held fixed, and score how many neighbor-relations come out matching the stored pattern — averaged over five trials). Fray damages the memory: more than half the pattern's bonds are cut down to a thread (K ≈ 0.03, far below anything functional). Probe again — recall drops. Then Heal: the network simply sits in its own pattern — settled, gently noisy, the learning rule running. No teacher, no stored copy, no error signal. Each frayed bond finds its endpoints already held in alignment by the surviving majority, and the rule does the only thing it knows: alignment above threshold, coupling grows. Probe once more.
In our verification runs the numbers were stark: recall 0.94 before damage, 0.87–0.90 frayed, and after one sleep episode — 1.00, with every single frayed bond regrown, on every seed we tried (102/102, 99/99, 104/104). The healed memory is sharper than the hand-carved original, because the rule also finishes the carving we did crudely: boundary bonds that fight the pattern get cleaned out.
Now the dark twin of the experiment. Vaporize destroys the same bonds completely — K exactly zero, not a thread. Heal all you like: nothing regrows. In our runs, zero of 102 vaporized bonds ever came back, and recall stayed stuck near 0.78. The reason is right there in the rule: K̇ = η[R₀(K)cos(Δθ) − 2K/r], and R₀(0) = 0 — K = 0 is an absorbing state. A bond with no coupling at all sees perfect alignment and does nothing about it, forever. The Shannon channel can strengthen a whisper into a voice, but it cannot speak first. Fraying leaves a whisper; vaporizing leaves silence. (This is the same lesson as the portrait's cold starts, now with stakes: growth needs a seed, and in a lattice the seed plus the neighbors' grip is enough.)
One more character from the learning-rule chapter has been waiting for its demonstration: the Fiedler channel — the structural term that pumps coupling into the network's bottlenecks. Here is the situation it exists for. Two tight communities, one bridge between them. The communities drift at different natural rates, so the bridge's endpoints cannot hold alignment — not because the bridge is unimportant, but because of where it stands. The Shannon channel, which sees only alignment, reads the bridge as a failing bond and executes it. Watch what happens to the network: the moment the bridge dies, λ₂ — the algebraic connectivity you met in the eigenvalues module — collapses to zero. One network becomes two. Structural richness is gone, and with it the capital C.
Toggle the structural channel on. It scores every bond not by alignment but by where it stands — the Fiedler sensitivity (v₂i − v₂j)², enormous exactly at bridges — and adds coupling there, paid for by a matching debit spread across the redundant bulk. In our runs the difference is total: Shannon-only, the bridge is dead and λ₂ = 0 within the first stretch of the run; with the structural channel, the bridge holds at K ≈ 3 and λ₂ stays up, indefinitely. And note one more thing the toggle shows: flip it on after the bridge has died, and the bridge comes back. The structural term is additive — it does not need a seed. What alignment cannot resurrect, structure can.
Step back and the module's three sections say one thing. A frozen wiring is a snapshot with an expiry date. The rule that moves the wiring has fates — birth, death, bistability — you can draw. And a network living under that rule is qualitatively unlike its frozen twin: damage it, and where a quenched system stays broken, the living one notices — because its memory and its dynamics were never separate things. The wiring is the memory. The dynamics are the repair crew. That, more than any single result, is what this site means by a living lattice.