The formal model in full: six theorems, written out

Work through the actual formalism of the source paper — the holon tuple, the learnable Topology-2-prime interface, the composition operator, and all six theorems with their proof sketches and exact conditions — so you can read the claims precisely instead of through a summary.

Track

Research

Module

5 · Research case studies

Prerequisites

When do many agents become one? Holonic composition · What makes a group genuinely emergent? The topology jump

A summary tells you a theorem exists; it does not let you check whether an architecture in front of you satisfies the theorem's hypotheses. For that you need the actual objects — what counts as a holon, what "mesh" means precisely, what the interface's type signature is. This lesson is deliberately dense: it is the reference you come back to when a vendor says "this composes correctly" and you want to ask "under which of the three conditions, and have you shown confluence of your substrate join?"

The recursion bottoms out at a single Belief–Desire–Intention agent and builds holons on top of it. Both are tuples. The agent carries the standard rationality constraints; the holon adds the machinery that makes a group behave like one thing.

the two tuples (ASCII rendering)
BDI agent:   a = < B, D, I, plan >
  B, D  subset of Phi       # beliefs, desires (formulas in a modal language)
  I     subset of D         # intentions
  constraints:
    internal consistency:   I union B  does NOT entail FALSE
    means-end coherence:    every intention in I has a plan in Plans
                            the agent believes achieves it
  plan: I -> Plans          # the planning function

Holon:   h = < A_h, compose_h, Sigma_h, phi_h, M_h >
  A_h        constituent agents or sub-holons (recursive; base = a BDI agent)
  compose_h  composition operator -> the holon's intention set I_h
  Sigma_h    SUBSTRATE: a shared information space for all members of A_h
  phi_h      the neuro-symbolic interface (see next section)
  M_h        trace memory: past interface invocations + outcomes

Well-defined holon intention set I_h

Three requirements at once: (1) internally consistent — I_h union B_h does not entail FALSE; (2) means–end coherent — every holon intention has a believable plan; (3) uniquely determined by the sub-holon intention sets together with the substrate dynamics. "Uniquely determined" is the clause people forget: a group whose collective intention depends on the order updates happened to arrive is NOT well-defined.

Substrate (kept abstract)

The paper does not fix the substrate to a blackboard, tuple-space, event log, or cache. The theorems depend only on three structural properties: write-visibility, ordering, and update atomicity. Any implementation with those properties is a faithful instantiation.

The distinguishing move of the paper is to give the symbolic–neural boundary a type and make it an object of learning. Ordinary Kautz Type 2 has a symbolic core that calls neural modules, but the call policy is wired at design time. Topology 2-prime promotes that policy to a function the agent updates from experience.

Topology 2-prime (ASCII rendering of the definition)
A holon is Topology 2-prime if there exist:

  symbolic component   S_h = < B_h, D_h, I_h >
  neural modules       N_h = { N_0, N_1, ..., N_k }
                       each N_j : X_j -> Y_j   (its own input/output spaces)
                       N_0 = the NULL module ("no invocation")
  argument space       Args = (disjoint union of the X_j) union { BOTTOM }
                       BOTTOM is paired with N_0

  such that the interface map has the typed signature

      phi_h :  S_h  x  Sigma_h  ->  N_h  x  Args

  i.e. given the symbolic state and the substrate state, phi_h decides
    (a) WHICH neural module to invoke, and
    (b) WITH WHAT argument;
  "no invocation" is the pair (N_0, BOTTOM).

  LEARNABLE: phi_h is updated as a function of the trace memory M_h.

  Recover vanilla Kautz Type 2 as the special case:
      M_h = empty   AND   phi_h constant (fixed at design time).

Read the signature slowly. The interface is not "the neural network"; it is the routing decision that sits in front of a bank of neural modules, including an explicit option to call nothing. That null module is load-bearing later: the emergence proof turns on sub-holons whose interfaces individually choose the null module on certain inputs, while the composite does not.

Composing two holons joins their substrates, aggregates their intention sets, lifts their interfaces, and unions their trace memories. The subtle parts are the substrate join (which must be confluent) and the meshing relation on intentions (which decides whether the aggregation is even coherent).

composition and meshing (ASCII rendering)
Composition  h = h1 compose h2 :
  substrate:   Sigma_h = Sigma_1  JOIN  Sigma_2
               under a unification relation ~ over shared identifiers
  intentions:  I_h = aggregate(I_1, I_2)      # the meshed intention set
  interface:   phi_h = the lifted interface on the joint domain
  memory:      M_h = M_1 union M_2 union M_(composite)

Meshing   I_1 mesh I_2   holds iff
  there exists a JOINT PLAN  p*  extending both p_1 and p_2  such that
    I_1 union I_2 union B_1 union B_2   does NOT entail FALSE
  AND
    no sub-goal of one agent's plan DEFEATS a sub-goal of the other's
    under the combined beliefs B_1 union B_2.

Confluent unification (~)

When the two substrates are merged, a shared identifier (say "document #4812") must resolve to the same object regardless of the order of merge steps. Confluence is exactly what guarantees the composite is uniquely determined rather than history-dependent — it is condition C3 of the Composition Theorem.

Defeat (between sub-goals)

One plan's sub-goal defeats another's when achieving the first makes the second unachievable under the shared beliefs. Meshing forbids this: cooperative plans may differ, but they may not sabotage each other.

This is the hinge result. It is an "if and only if", which is stronger than most composition claims in the multi-agent literature: it says the three conditions are not merely sufficient recipes but genuinely necessary.

T1 statement (ASCII rendering)
T1 (Composition).  Let h1, h2 be Topology 2-prime holons.
The composition h = h1 compose h2 admits a WELL-DEFINED holon-level
intention set I_h  IF AND ONLY IF all three hold:

  C1  Meshing:              I_1 mesh I_2.
  C2  Mutual responsiveness: for each i, there is a write-read channel
                             through Sigma_h by which h_i can observe
                             changes to the OTHER's intentions within
                             bounded latency.
  C3  Substrate compat.:    the unification ~ on Sigma_1 JOIN Sigma_2 is
                             well-defined and confluent.

1. 1Forward direction (well-defined => C1,C2,C3). Suppose I_h is well-defined. Internal consistency plus means-end coherence of I_h force C1: if the sub-intentions did not mesh, no joint plan exists and I_h cannot be means-end coherent. Drop C2: the substrate cannot reflect one sub-holon's updates to the other, so the "joint" plan admits divergent executions — this breaks the mutual-responsiveness clause of shared cooperative activity and contradicts well-definedness. Drop C3: the substrate join is ambiguous, so the aggregated intention set is not unique, contradicting the "uniquely determined" clause.
2. 2Backward direction (C1,C2,C3 => well-defined). Construct I_h explicitly as the closure, under the composition operator, of the meshed intentions, with any conflicts resolved by the substrate's confluent unification. Internal consistency follows from C1; means-end coherence of the joint plan from C1 together with C2; uniqueness from C3.
3. 3Read-off for engineers (the paper's Remark): an architecture failing any one of C1–C3 can still LOOK coherent under a lucky event ordering while admitting observation-dependent inconsistencies under an unlucky one — a defect that is sequence-dependent and therefore easy to miss in testing.

T2 governs how a sub-holon intention appears in its parent; T6 generalises T1 to the realistic case where sub-holons conflict locally but can still cooperate under arbitration. T1 turns out to be the meshing limit of T6.

T2 and T6 (ASCII rendering)
T2 (Hierarchical Persistence).  Let h be a sub-holon of H and let
  ii be an intention of h.  Then there is a LIFTED intention  lift(ii)
  in I_H such that:
  (a) lift(ii) is consistent with all other lifted sub-holon intentions;
  (b) reconsidering lift(ii) at level H triggers reconsidering ii at
      level h ONLY IF the parent policy propagates the event downward;
  (c) abandoning ii at level h leaves lift(ii) PROVISIONALLY SATISFIED
      at level H until the parent policy revisits it.
  Proof idea: lift(.) is the canonical map induced by the parent's
  intention aggregator; uniqueness of the image follows, consistency
  follows from T1 applied to H; (b)/(c) encode resource-bounded
  rationality — levels do not spontaneously churn each other.

Tribunal  T_h : I_1 x I_2 x (beliefs) -> (intentions)   must satisfy
  Conservativity: if I_1 mesh I_2 then T_h returns I_1 union I_2
  Consistency:    output union beliefs does NOT entail FALSE
  Maximality:     output is a MAXIMAL consistent subset of I_1 union I_2
                  (remove the FEWEST intentions needed to restore consistency)

T6 (Arbitrated Composition).  If I_1, I_2 need not mesh, T_h is a
  tribunal with the three properties above, and C2 and C3 of T1 hold,
  then the arbitrated composition admits a well-defined holon intention
  set  I* = T_h(I_1, I_2, B_1 union B_2).
  Corollary: when I_1 mesh I_2, conservativity gives I* = I_1 union I_2,
  so T1 is exactly the meshing limit of T6.

Proof sketch of T6, in words: consistency of the tribunal gives internal consistency of I*; means-end coherence holds because each retained intention keeps a plan from one of the two sub-holons (and any substituted intention inherits a plan from the tribunal's output); mutual responsiveness (C2) and substrate compatibility (C3) are inherited from the underlying composition because the tribunal operates on intention sets, not on substrate structure. Status: PROVEN (proof sketch). The paper leaves the concrete arbitration policy — priorities, voting, learned arbitration, contract-based reasoning — open; only the three properties are fixed.

These are the emergence results, proved along the coupling-tightness axis of Kautz's taxonomy (Type 2 looser than Type 5 looser than Type 6). T3 is an existence theorem with an explicit witness built from the null-module trick; T4 shows the top of the axis is absorbing.

T3 witness and T4 (ASCII rendering)
T3 (Emergence by Topology Jump).  There exist holons h = h1 compose h2
  with h1, h2 of Topology 2-prime but h of Kautz Type 5 or higher.

  Witness construction:
    let N_1, N_2 produce structured outputs in a shared space Z,
    with a CROSS-MODAL subset Z_x (elements mixing both modalities).
    Each sub-holon interface DEFERS on cross-modal input:
        phi_i(s, sigma) = (N_0, BOTTOM)   whenever sigma encodes a Z_x element
    Compose, and define the LIFTED interface:
        phi_h(s, sigma) =
          (N_x, a_x)       if sigma encodes a Z_x element involving
                           outputs of BOTH N_1 and N_2
          phi_i(s, sigma)  otherwise
    where N_x is a designated JOINT-CONDITIONING module.
  => phi_h conditions on the JOINT neural state of both parts on at
     least one input, while neither phi_1 nor phi_2 does individually.
  => that joint conditioning is the defining property of Type 5.
  => h sits at strictly tighter coupling than any sub-holon: a JUMP.

  Corollary (Irreducibility): there is a property P with h |= P but
    h_i does NOT satisfy P for every sub-holon h_i.

T4 (Topology 6 Closure).  If h is Kautz Type 6 and h compose h'
  satisfies C1-C3, then h compose h' is also Type 6.
  Proof idea: Type 6 means the symbolic operators are NOT a separable
  component but patterns inside the neural substrate — there is no
  extraction map faithful to the symbolic semantics. Any symbolic
  decision in the composite that touches h's substrate is mediated by
  those neural-embedded operators, so the composite has no separable
  symbolic core either.
  Corollary (Irreversibility): a Type 6 holon has no lossless
  decomposition into Type <= 5 parts that recompose into it.

T5 is the quantitative result. It rests on typing the event stream by class, attaching a frequency and a relevance density to each class, and scoring a reconsideration policy by its per-class true-positive and false-positive rates. The payoff is a clean cost expression, an exponential concentration bound, and the informed-vs-blind separation.

T5 ingredients and claims (ASCII rendering)
Event space typed by class:  E = disjoint union over classes c of E_c
  class prior     p_c   = Pr[event in class c],  sum_c p_c = 1
  relevance dens. rho_c = Pr[event of class c actually requires
                             reconsideration to keep coherence]
Class-typed policy  Pi = (Pi_c), each Pi_c acts only on class c, with
  true-positive rate  alpha_c = Pr[trigger | class c, relevant]
  false-positive rate beta_c  = Pr[trigger | class c, NOT relevant]
  marginal selectivity eps_c  = rho_c*alpha_c + (1-rho_c)*beta_c
  RELEVANCE-BLIND:    alpha_c = beta_c for all c
  RELEVANCE-INFORMED: gap alpha_c - beta_c > 0 for some c

Let  R = sum_c p_c*rho_c   (aggregate relevance),  beta_max = max_c beta_c.

(i)  Cost:  E[cost] = C_rec * sum_c p_c * [ rho_c*alpha_c + (1-rho_c)*beta_c ]
(ii) Concentration:  for an i.i.d. window of size n and any D>0,
       Pr[ |empirical_miss - mu| >= D ]  <=  2*exp(-2*n*D^2)
     where mu = sum_c p_c*rho_c*(1-alpha_c).
(iii)Coherence:  if alpha_c >= 1 - delta/(|C|*p_c*rho_c) for every c,
       then miss rate mu <= delta  (coherence preserved within delta).
(iv) Informed-blind separation:
       informed (alpha_c = 1):  miss = 0 at cost
           <= C_rec * ( R + (1-R)*beta_max )
       any blind policy reaching miss <= delta costs
           >= C_rec * (R - delta) / rho_max,   rho_max = max_c rho_c
       in the homogeneous regime rho_c = R this is >= C_rec*(1 - delta/R),
       and cost_blind / cost_informed -> UNBOUNDED as delta -> 0
       (whenever R + (1-R)*beta_max < 1).

Proof sketch, in words: (i) the conditional trigger probability per class is the relevance-weighted mix of TPR and FPR; average over classes and multiply by the per-reconsideration cost. (ii) the per-event "relevant but missed" indicator is bounded in [0,1] with the stated mean, so Hoeffding's inequality over n independent draws gives the exponential tail. (iii) choosing each class's TPR high enough caps that class's contribution to the miss rate at delta/|C|; summing gives miss <= delta. (iv) set TPR = 1 for the informed bound; for the blind bound use alpha_c = beta_c and bound the achievable miss rate below, then rearrange. Status: PROVEN (proof sketch). A supporting proposition shows that if the sensor is sound, the empirical class priors and relevance densities converge to their true values, so the quantities in T5 are estimable from traces.

Put the pieces together on one synthetic system: a reconciliation holon built from an invoice agent and a manifest agent, run under a typed-event policy.

synthetic — full pass through the formalism
STEP 1  Base agents (BDI):
  invoice agent:  I_1 = { approve_if_matches }
  manifest agent: I_2 = { hold_if_quantity_mismatch }

STEP 2  Check T1's conditions for composing them:
  C1 mesh?   joint plan where approve and hold do not defeat each other
             exists IF "approve" is gated on "no mismatch". -> holds
  C2 resp.?  invoice agent can see a hold via Sigma within bounded
             latency. -> provision a write-read channel
  C3 conf.?  "invoice #4812" and "manifest for #4812" unify to one
             object confluently. -> define the unification rule
  => T1 satisfied: the holon has a well-defined collective intention.

STEP 3  Topology check (T3):
  each agent DEFERS (null module) on cross-document features;
  the lifted interface fires a JOINT module on (invoice_total,
  manifest_qty) together -> Type 5 coupling -> topology jump.

STEP 4  Event policy (T5): classes = {arrival, deadline, exception}
  learn high TPR on deadline/exception (high rho_c),
  low trigger on routine arrival (low rho_c)
  -> informed policy buys low miss rate cheaply (separation (iv)).

STEP 5  If a third, conflicting agent is added and meshing fails:
  invoke a tribunal (T6) that is conservative + consistent + maximal
  -> still a well-defined holon intention set I* = T_h(...).

Notice every step is a check you can actually perform at design time: does a joint plan exist, is there a responsiveness channel, is the unification confluent, does the lifted interface condition jointly, is the policy class-typed. The formalism's value is turning "will this multi-agent system cohere?" into a finite checklist with named failure modes.

• ▸PROVEN (deductive, proof sketches): all six theorems (T1–T6) and the sensor-soundness proposition. These are results about the model.
• ▸ASSUMPTIONS baked into the model: finite intention sets; a global (stationary) event distribution; the interface treated as an abstract learnable function rather than a concrete algorithm.
• ▸NOT PROVIDED here: a learning-theoretic analysis of the interface (sample complexity, regret). The paper flags this as future work.
• ▸NOT TESTED: none of these theorems is empirically validated in the paper; that is the subject of case lesson 5.
• ▸OPEN: closure of Type 5 under composition; the spectrum of attainable topology jumps as a function of sub-holon diversity.

Expected direction: "Uniquely determined" means the aggregated intention set does not depend on the order in which substrate updates are merged. C3 requires the unification relation on the joined substrate to be confluent — same result regardless of merge order — which is exactly what makes the aggregate unique. If C3 failed while C1/C2 held, the sub-intentions would still mesh and the agents could still observe each other, but a shared identifier could resolve to different objects depending on update ordering, so the same inputs could yield different collective intentions across runs: an order-dependent / history-dependent, non-reproducible holon intention.

What is the type signature of the Topology-2-prime interface, and what does the null module encode?

phi_h : S_h x Sigma_h -> N_h x Args — given symbolic state and substrate state it picks which neural module to invoke and with what argument. The null module N_0 (paired with BOTTOM) encodes "no invocation," the explicit choice to call nothing.

In the T1 proof, what goes wrong if C2 (mutual responsiveness) is dropped?

The substrate cannot reflect one sub-holon's intention updates to the other, so the joint plan admits divergent executions, breaking the mutual-responsiveness requirement of shared cooperative activity and contradicting well-definedness.

How is T1 related to T6?

T1 is the meshing limit of T6: by the tribunal's conservativity property, when the sub-holon intentions mesh the tribunal returns their union, recovering exactly the T1 composition.

State the informed-vs-blind separation of T5(iv) in terms of cost as the miss-rate target shrinks.

An informed policy (TPR=1) reaches zero miss at cost bounded by C_rec(R + (1-R)*beta_max); a blind policy reaching miss <= delta costs at least C_rec(R-delta)/rho_max, so the ratio cost_blind/cost_informed grows without bound as delta -> 0 (when R + (1-R)*beta_max < 1).