The Spiking Prototype: a Nambu-Goto-Enforced Substrate and What It Does (Not) Achieve
v0.1 — a 3-D LIF/ALIF blob with node-placed functional plasticity, two slow homeostats, Dale's-law E/I balance, and a global surface-minimisation watcher; the honest baseline the local-rule programme must beat
Criticality and Nambu-Goto surface minimisation are emergent global properties: neither may be enforced from outside at runtime without making the result circular. This paper presents v0.1, a deliberately honest prototype that enforces the geometry anyway — a global watcher slides the free interior of a 3-D embedding down the gradient of the physical-network surface action S = Σ π w L and rewires by proximity — so there is a working substrate and a measured baseline. The atomic unit is a leaky integrate-and-fire neuron with spike-frequency adaptation; its functional plasticity lives on the node (a learnable membrane time constant and threshold), not on O(E) edges, and its synapses are scalar weights under local STDP behind a three-factor interface. Excitability is regulated by two slow, literature-grounded homeostats — Triesch intrinsic plasticity (threshold) and Turrigiano synaptic scaling (weight) — over a Dale's-law excitatory/inhibitory balance. We justify each choice and then test the substrate honestly. Across 36 runs (sizes, seeds, and ablation arms) the watcher minimises surface every pass; the population is alive and rate-regulated; and the branching ratio σ_MR sits sub-critical at ≈0.32, tight across seeds — the expected negative under globally-enforced geometry. The load-bearing finding is a clean discriminator: ablate the homeostat and σ_MR collapses to ≈1.0 (synchrony), so excitability regulation, not the geometry, is what keeps the substrate measurable. v0.1's job is to establish the substrate and prove its central property is not an artefact of its regulator — and to make the case that the enforced geometry must instead be *earned* by a local rule, which is the next paper.
Introduction
Purpose of this section. State the reframe, the honest scope of v0.1, and the falsifiable question it sets up.
- The reframe. Criticality and Nambu-Goto surface minimisation are emergent global properties of a population of locally-acting units. Enforcing either at runtime — a global servo that pins mean activity, or a watcher that solves the surface PDE and moves nodes — makes "it emerged" circular: the consequence has been hardcoded as the cause.
- What v0.1 is. A deliberately honest prototype that does enforce the geometry (a global watcher), so there is a working substrate and a measured baseline. The watcher is named as a shortcut, not a result.
- What v0.1 is for. To (a) justify a minimal spiking substrate component by component, (b) characterise honestly what enforced-NG buys — expected: not criticality — and (c) state the question the next paper answers: can a purely local rule earn the surface minimisation the watcher fakes?
- Predecessor. SORN (self-organising recurrent networks; Lazar–Pipa–Triesch) is the closest ancestor — local plasticity + homeostasis producing structured dynamics — which we extend with an explicit 3-D physical geometry.
The architecture, and why each piece
Purpose of this section. Justify the atomic unit and every component — the spine of the paper. Each choice is defended against its alternative.
- The atomic unit — LIF/ALIF spiking, not a continuous map. Why spiking over the earlier continuous coupled-map-lattice with KAN edges: O(N) state vs O(E·(G+k)) splines, biologically faithful, async-portable (the same unit runs clock-driven now and event-driven later), and — decisively — spikes make avalanche/criticality measurement native, which a dense continuous field does not (its branching-ratio estimator degenerates). Spike-frequency adaptation (ALIF) gives a node-local slow timescale.[3][4]
- Functional plasticity on the node, not the edge. A learnable membrane time constant (parametric-LIF) and threshold per neuron — O(N) — instead of a learnable function per synapse — O(E). The synapse is ~a scalar gain; the nonlinearity lives in the cell.
- Scalar synapses + STDP behind a three-factor interface. Local, online, and structured so a top-down learning signal (e-prop) drops in unchanged at scale — the sync→async bridge.[5]
- Two slow homeostats — both legs. Triesch intrinsic plasticity slides the threshold toward a firing-rate set-point (regulates rate); Turrigiano synaptic scaling multiplicatively rescales incoming excitatory weights (regulates magnitude, bounds STDP runaway). Each reads only the cell's own activity — never a population aggregate.[6][7]
- Dale's-law E/I balance. A fixed inhibitory subset desynchronises the recurrent net; without it the population synchronises and presents manufactured criticality (σ_MR→1). The fast lateral divisive-normalisation leg is validated but kept off the default path.[8][9]
- 3-D embedding + the global NG watcher. Each cell carries a position in ℝ³; the watcher slides the free interior down the gradient of the physical-network surface action S = Σ_e π w L_e and rewires by proximity, so geometry feeds back onto the dynamics. This is the enforced shortcut — stated plainly as the thing 0002 replaces.[10]
What we test, and why
Purpose of this section. Lay out the probe suite and the ablation matrix, and justify each measurement — what question it answers and what would falsify it.
- NG-is-happening witness. Confirm the watcher actually minimises surface: the per-pass descent ΔS ≤ 0 (a backtracking line search guarantees it at every scale). The honest nuance: the proximity rewire densifies the graph, so total S can rise even as each node descends — the witness is the descent, not total S.
- Criticality battery. The central question — does criticality emerge? Gauges: the Wilting–Priesemann branching ratio σ_MR (subsampling-robust), avalanche size/duration power-law exponents and the crackling relation, Lempel-Ziv complexity (rate-controlled), and the 1/f aperiodic exponent. Expectation: sub-critical under globally-enforced geometry — the honest negative, and the baseline 0002 must beat.[11][12]
- Perturbation response. Poke one neuron and measure how the perturbation propagates against an identical twin (spike-difference spread, divergence trend, perturbational complexity / PCI-style LZc) — does the substrate support non-trivial propagation?[13]
- Ablations as discriminators — how we attribute behaviour to components. Each arm removes one axis: NG-off (geometry frozen — is the geometry doing anything?), homeostat-off (excitability regulation — the load-bearing test), STDP-off (synapses frozen), scaling-off (does Turrigiano bound runaway?), lateral-on (the fast divisive-normalisation leg). The contrast between arms, not any single number, is what isolates a mechanism's role.
Results
Purpose of this section. Report the measured baseline — 36 runs (sizes 4³ and
5³, seeds 0–2, six ablation arms) — and read the honest negatives and the one
load-bearing positive. All numbers are from the certified sweep
(verdict.json: prototype_works, ng_minimisation, discriminator_present all
true).
The substrate is alive and the watcher minimises surface. Every NG-on arm
descends surface on every pass (frac_passes_descending = 1.0); the migration
removes surface scaling with the free interior (≈0.4 at side 4, ≈2.4–4.0 at side 5);
the population holds a regulated firing rate (~0.06/step).
Criticality is sub-critical — the expected negative. The branching ratio on the full operating arm is tight across seeds and sizes:
| size | seed 0 | seed 1 | seed 2 |
|---|---|---|---|
| 4³ (N=64) | 0.234 | 0.299 | 0.308 |
| 5³ (N=125) | 0.322 | 0.322 | 0.326 |
σ_MR ≈ 0.32 (median), well below 1 — under globally-enforced geometry the substrate does not self-organise to criticality. A critical, well-posed σ_MR≈1 would have been the surprise.
The load-bearing result — the homeostat, not the geometry, keeps it measurable. Ablating the slow excitability axis collapses σ_MR to ≈1.0 (median 1.0000 vs the full arm's 0.32) — synchronised degeneracy, not criticality. This is the cleanest discriminator in the matrix and the figure below.
Perturbations propagate. A single poke spreads through the population with perturbational complexity (LZc of the spike-difference raster) of 0.95–0.99 across all full-arm runs — non-trivial propagation despite sub-critical branching.
A real scale effect, reported as data. At size 5³ the control arms
(homeostat-off, NG-off, STDP-off) fall below the avalanche-applicability floor — they
lack a leg the full configuration has, so they degrade as designed. This is recorded
(arms_not_applicable), not hidden by the gate.
A baseline, not the thing
Purpose of this section. Close by stating plainly why this substrate is a baseline, and hand off the falsifiable question to the next paper.
- The circularity, named. The watcher solves the surface problem globally and moves the nodes there. So "the substrate minimises its wiring surface" is true by construction, not by emergence — we hand-solved the PDE and called it a property. v0.1 is honest about this: it is the control, not the claim.
- What the baseline establishes. A justified spiking substrate; a measured, reproducible operating point (σ_MR≈0.32, sub-critical); and a clean attribution — the homeostat is load-bearing, the enforced geometry is not what makes the dynamics measurable.
- The question for 0002. Can a purely local update rule — each cell reading only its own state and its immediate neighbours, never the global action — be distilled such that, run with the watcher off, the wiring surface still descends? And does the same frozen rule hold across orders of magnitude in N (a law, not a fit)? That is the contribution of the next paper.
- Deferred. The tissue comparison and the holographic area-law test need a second timescale and the scaled async substrate; they belong to v0.2/v0.3, not here.