What makes a group genuinely emergent? The topology jump

Understand how a source paper turns the slippery word "emergence" into a decidable structural property: a composed holon can sit at a strictly tighter neuro-symbolic coupling than any of its parts, and that jump is readable off the interfaces — not inferred from behaviour.

Track

Research

Module

5 · Research case studies

Prerequisites

When do many agents become one? Holonic composition · Deterministic vs probabilistic agents

People say a flock, a market, or a multi-agent LLM system "exhibits emergent behaviour." The trouble is that the claim is almost always qualitative: you watch the system, something surprising happens, and you call it emergence. That is unfalsifiable and tells a designer nothing. The source paper asks for an emergence indicator that is a property of the composition itself — decidable from the structure, before you run anything.

Its answer uses the coupling dimension of Kautz's neuro-symbolic taxonomy. Think of a spectrum of how tightly a system's symbolic reasoning and neural computation are entangled. At a looser coupling, a symbolic core calls neural modules as subroutines and the two stay individuated (the paper's learnable version of this is Topology 2-prime, from case lesson 1). Tighter, the symbolic decision is jointly conditioned on neural and symbolic state at once (Kautz Type 5). Tightest, the symbolic operators are no longer a separable component at all — they live as patterns inside the neural substrate (Type 6). Emergence, the paper argues, is a composition landing higher on this coupling axis than any of its parts.

Recall the setup from case lesson 1: a holon has sub-holons, a shared substrate, a learnable neuro-symbolic interface, and a trace memory. Each sub-holon's interface decides which neural module to invoke and with what argument. The key structural fact the paper exploits is how those interfaces behave on inputs that mix the modalities of two different sub-holons — "cross-modal" joint states that only make sense once the two are put together.

Coupling topology (the relevant axis)

Type 2 / 2-prime: symbolic core invokes neural modules as subroutines, cores individuated. Type 5: symbolic decisions jointly conditioned on neural and symbolic state together, cores still separable. Type 6: symbolic operators realised inside the neural substrate, with no faithful extraction back to a separable symbolic core.

Topology jump

A composition whose coupling topology is strictly tighter than that of every one of its sub-holons.

Irreducibility

The existence of a property that the holon has but no individual sub-holon has — the structural counterpart of "more than the sum of the parts."

Lifted intention

The image of a sub-holon's intention inside the parent holon's intention set, under the map induced by the composition operator.

The paper's emergence result (it labels it T3) is an existence theorem proved by exhibiting an explicit witness. Take two sub-holons that each, on their own, refuse to condition their symbolic decisions on cross-modal joint state — individually each one just defers (invokes nothing) when it sees a mixed-modality input. Now compose them and define the lifted interface so that, when the input genuinely mixes outputs of both, a joint-conditioning module fires. The composite now makes decisions conditioned on the joint neural state of both parts, which by definition is the tighter Type-5 coupling — even though neither part did so alone. The composite therefore sits at a strictly higher coupling topology: a topology jump, and a concrete witness of irreducibility.

A complementary result (it labels it T4, Topology 6 Closure) handles the top of the coupling scale. If one of the parts is already Type 6 — its symbolic operators are inseparable from its neural substrate — then any well-defined composition with it is again Type 6: you cannot compose your way back down to a separable symbolic core. Full neuro-symbolic integration, once reached, is absorbing. Put together, the two results impose a partial monotonicity on the hierarchy along the coupling axis: tightness can rise under composition (T3) and cannot fall once it hits the top (T4).

synthetic — the witness, in words
sub-holon A:  on a cross-modal input, interface -> (null module)   # defers
sub-holon B:  on a cross-modal input, interface -> (null module)   # defers

compose(A, B):
  on an input mixing A's and B's outputs:
     lifted interface -> JOINT-CONDITIONING module   # fires on joint state
  => composite conditions on joint neural state of BOTH parts
  => coupling is strictly tighter (Type 5) than either part (Type 2-prime)
  => "topology jump"

Status: PROVEN (proof sketches given). T3 is an existence claim — it shows such holons exist, not that every composition jumps. The paper is explicit that whether the middle of the scale (Type 5) is itself closed under composition is an OPEN question, and that the vanilla fixed-interface Type 2 case does not get the jump for free: the degrees of freedom that let the lifted interface absorb cross-modal joint state come from the interface being learnable. Emergence here is a property of the learnable interface, not of holonic composition on its own.

Picture two LLM-backed agents. Agent A reads invoices; agent B reads shipping manifests. Separately, each handles only its own document type and ignores signals that only make sense across both. Compose them into a reconciliation holon, and the learned joint interface starts conditioning on features that exist only in the pairing — a quantity on the invoice set against a quantity on the manifest.

synthetic — reconciliation holon
A alone:  invoice total extracted            (no manifest context)
B alone:  manifest quantity extracted         (no invoice context)

holon:    decision conditioned on (invoice total, manifest quantity)
          TOGETHER -> flags a mismatch neither agent could see alone

The mismatch-detection capability is irreducible: it is a property of the holon that is provably not a property of either agent. That is the paper's sharpened notion of emergence — and notice it is claimed structurally, by pointing at the joint-conditioning interface, not by admiring the output after the fact.

For the hierarchy to be usable, intentions must behave sensibly as you move up and down it. The paper's Hierarchical Persistence result (it labels it T2) says every sub-holon intention has a unique lifted image in the parent, that image is consistent with the other lifted intentions, and reconsideration is asymmetric: the parent only revisits a lower intention when it propagates a triggering event downward, and a sub-holon abandoning a goal leaves the lifted version provisionally satisfied above until the parent revisits it. This is the resource-bounded-rationality idea made structural — a higher level does not spontaneously churn over lower-level details, and vice versa.

Why a security-minded engineer should care: a provable topology jump is a principled way to claim genuine emergence in systems that combine LLMs and symbolic reasoning — instead of the usual marketing use of the word. But the same irreducibility cuts both ways. A capability that exists only at the holon level also means a failure or a manipulation can exist only at the holon level: cross-agent prompt injection and agent collusion are exactly attacks on the joint behaviour that no single agent's logs will show. If emergence is real and structural, so is emergent risk.

Honest status of this lesson's claims:

• ▸PROVEN (deductively, with proof sketches): the Emergence Theorem (there exist holons at strictly tighter coupling than any sub-holon), the Topology 6 Closure result, and Hierarchical Persistence.
• ▸SCOPE: the "higher topology" ordering is only along the coupling-tightness axis (2 to 5 to 6). It is NOT a total order over all six Kautz classes; Types 1, 3, 4 are categorically distinct and outside these theorems.
• ▸DEPENDENCY: the topology jump relies on the interface being learnable (Topology 2-prime). With a fixed design-time interface the jump does not come for free — the paper is explicit about this.
• ▸OPEN: whether the middle coupling class (Type 5) is itself closed under composition; and the full spectrum of attainable jumps as a function of how diverse the sub-holons are.
• ▸NOT YET MEASURED: these are structural/decidable claims about the model. Detecting a predicted cross-modal conditioning in a real deployment is left to the interventional experiments of case lesson 3.

Expected direction: They would have to show a topology jump: the composite conditions its symbolic decisions on cross-modal joint state of both parts (tighter Kautz coupling, e.g. Type 5) when neither part does so individually — a property readable from the lifted interface, not from watching outputs. Limitation: the "higher topology" ordering is only along the coupling-tightness axis (2 to 5 to 6); it is not a linear ranking over all six Kautz classes, and Types 1/3/4 are outside these claims. Also, the jump depends on the interface being learnable.

What does the paper use as its decidable indicator of emergence?

A topology jump: the composed holon sits at strictly tighter neuro-symbolic coupling (higher Kautz topology along the coupling axis) than any of its sub-holons, readable from the interfaces.

Is the Kautz ordering used a total order over all six types?

No. It is only along the coupling-tightness axis (Type 2 looser than 5 looser than 6). Types 1, 3, and 4 are categorically distinct and the topology-jump theorems do not range over them.

What does the Topology 6 Closure result say?

If one part is already Type 6 (symbolic operators inseparable from the neural substrate), any well-defined composition with it is again Type 6 — full integration is absorbing; you cannot compose back to a separable symbolic core.

Does holonic composition alone produce the topology jump?

No — the jump depends on the interface being learnable (Topology 2-prime). With a fixed design-time interface the composite has no degrees of freedom to absorb cross-modal joint state, so the jump is a property of the learnable interface, not of composition by itself.