Emergence Theory¶

Jean-Paul Niko · February 2026

We develop a formal theory of emergence for cognitive assemblies---collections of cognitive engines (human, machine, or hybrid) whose joint capabilities exceed those of any individual component. Three classes of emergence are defined with increasing structural depth: quantitative emergence (\(E_1\)), where the assembly outperforms its best component in measurable capability; qualitative emergence (\(E_2\)), where the assembly activates capability types that no individual component possesses; and structural emergence (\(E_3\)), where the assembly's conceptual topos contains truth values absent from every component's topos. We define measures for each class---the emergence surplus \(\varepsilon_1\), the activation count \(\varepsilon_2\) and novelty multiplier \(\mu_t\), and the structural gap surplus \(\varepsilon_3\)---and combine them into a total emergence value \(V(A)\) that serves as the objective function for assembly optimization. We prove that the greedy assembly algorithm achieves a \((1-1/e)\)-approximation when \(V\) is submodular, establish conditions under which each emergence class satisfies or violates submodularity, and define the emergence tensor \(E_{st}\) that captures second-order interaction structure. Throughout, we connect the emergence framework to the compatibility matrix \(K\), the Conceptual Irreversibility Theorem, the ELO rating system, the attention simplex, and the hypervisor loss function from earlier parts of the paper.

%═══════════════════════════════════════════════════════════════════════════════

Motivation: The Missing Piece¶

%═══════════════════════════════════════════════════════════════════════════════

The Intelligence as Geometry framework provides three tools for analyzing multi-component cognitive systems:

[nosep] - Intelligence vectors \(\bI \in [0,\infty)^8\), measuring per-type capability (Part I). - The synergy factor \(\mathrm{Syn}(B) = \|\bI_B\| / \max_i \|\bI_i\|\), quantifying how much the bundle outperforms its best component (Part II). - Bundle intelligence \(\bI_B\), the resultant intelligence vector of an assembly computed via the compatibility matrix \(K\) (Part II).

What is missing is a theory of what new capabilities appear when components are assembled, how to measure them, what they are worth, and how to manage them. This is the emergence theory.

The three classes correspond to increasing levels of novelty:

\begin{center}

[Table — see PDF]

\end{center}

%═══════════════════════════════════════════════════════════════════════════════

Preliminaries¶

%═══════════════════════════════════════════════════════════════════════════════

We fix notation. Let \(T = \{1, \ldots, 8\}\) index the eight intelligence types (linguistic, spatial, social, symbolic, mnemonic, evaluative, auditory, kinesthetic). An assembly is a finite set \(A = \{e_1, \ldots, e_n\}\) of cognitive engines, each with intelligence vector \(\bI_i = (I_{i,1}, \ldots, I_{i,8}) \in [0,\infty)^8\), ELO vector \(E_i \in \\mathbb{R}^{n(e)}\), and per-token cost \(c_i > 0\).

The assembly's joint intelligence vector \(\bI_A\) is computed via the compatibility matrix \(K \in \R^{8 \times 8}\): \begin{keyeq} [ I_{A,t} = \sum_{i=1}^{n} I_{i,t} + \sum_{\substack{i < j s \neq t}} (K_{st} - 1) \cdot I_{i,s} \cdot I_{j,t} ] \end{keyeq} where the second sum captures pairwise cross-type synergy (Part II, with the acknowledged limitation that this is a pairwise approximation; see Remark ref:rem:pairwise).

The activation gate from Definition 4.2 uses threshold \(\theta = 0.1\) cog: type \(t\) is active in engine \(e_i\) if and only if \(I_{i,t} > \theta\). The set of active types for engine \(e_i\) is \(T_{\mathrm{active}}(e_i) = \{t \in T : I_{i,t} > \theta\}\).

The synergy formula uses pairwise interactions only. Higher-order terms (three-way synergy, etc.)\ are omitted as a first approximation. The emergence tensor (Section ref:sec:tensor) partially addresses this by capturing second-order structure, but genuine \(k\)-body interactions (\(k \geq 3\)) remain an open problem.

%═══════════════════════════════════════════════════════════════════════════════

Class \(E_1\): Quantitative Emergence¶

%═══════════════════════════════════════════════════════════════════════════════

\tA\; An assembly \(A\) exhibits quantitative emergence in type \(t\) when [ I_{A,t} > \max_{i} \, I_{i,t}. ] The assembly outperforms its best component in at least one intelligence type.

\tA\; The emergence surplus of assembly \(A\) in type \(t\) is \begin{keyeq} [ \varepsilon_1(A, t) \;=\; I_{A,t} - \max_{i} \, I_{i,t}. ] \end{keyeq} The surplus vector \(\varepsilon_1(A) = (\varepsilon_1(A,1), \ldots, \varepsilon_1(A,8)) \in \\mathbb{R}^{n(e)}\) captures the full emergence profile.

\tA\;

The emergence surplus in type \(t\) admits the first-order approximation [ \varepsilon_1(A, t) \;\approx\; \frac{1}{n} \sum_{\substack{i \neq j s \neq t}} (K_{st} - 1) \cdot I_{i,s} \cdot I_{j,t}. ] In particular, \(\varepsilon_1(A,t) > 0\) whenever there exist components \(e_i, e_j\) with \(I_{i,s} > 0\), \(I_{j,t} > 0\) for some pair \((s,t)\) with \(K_{st} > 1\).

Expand \(I_{A,t}\) using the synergy formula and subtract \(\max_i I_{i,t}\). The direct sum \(\sum_i I_{i,t}\) exceeds \(\max_i I_{i,t}\) by \(\sum_i I_{i,t} - \max_i I_{i,t} \geq 0\) (with equality only for single-component assemblies). The cross-type synergy terms contribute \((K_{st}-1) \cdot I_{i,s} \cdot I_{j,t}\) for each ordered pair; the \(1/n\) normalization converts from sum to per-component average, matching the extensive scaling of the synergy formula. When \(K_{st} > 1\), each such term is strictly positive.

\tA\; The scalar value of the quantitative emergence is [ V_1(A) = \sum_{t \in T} \varepsilon_1(A, t) \cdot p_t ] where \(p_t > 0\) is the market price of one cog of type-\(t\) intelligence (or the task-distribution-weighted relevance of type \(t\)).

Properties of \(V_1\)¶

\tA\;

The function \(V_1: 2^{\mathcal{E}} \to \R_{\geq 0}\), where \(\mathcal{E}\) is the catalog of available engines, is submodular: [ V_1(A \cup {e}) - V_1(A) \;\geq\; V_1(B \cup {e}) - V_1(B) \quad \text{for all } A \subseteq B,\; e \notin B. ] That is, the marginal value of adding a component decreases as the assembly grows.

The emergence surplus \(\varepsilon_1(A,t) = I_{A,t} - \max_i I_{i,t}\) is a sum of terms of the form \((K_{st}-1) \cdot I_{i,s} \cdot I_{j,t}\) (for \(K_{st} > 1\); negative-synergy pairs contribute negatively). Adding engine \(e\) to a smaller assembly \(A\) introduces cross-terms with every existing component. In the larger assembly \(B \supseteq A\), many of these cross-type interactions are already partially satisfied by additional components in \(B \setminus A\). Formally, the contribution of \(e\) to \(\varepsilon_1\) in type \(t\) includes terms \((K_{st}-1) \cdot I_{e,s} \cdot I_{j,t}\) for each \(j \in A\). Since \(A \subseteq B\), every such term appears in \(\Delta V_1(e \mid B)\) as well, but \(B\) also contains components whose presence means the max in the surplus definition is weakly larger, reducing the net gain. The \(\max\) operator is concave, and composition of a linear function with a concave function preserves submodularity.

%═══════════════════════════════════════════════════════════════════════════════

Class \(E_2\): Qualitative Emergence¶

%═══════════════════════════════════════════════════════════════════════════════

\tB\; An assembly \(A\) exhibits qualitative emergence in type \(t\) when \begin{keyeq} [ I_{A,t} > \theta \quad\text{and}\quad I_{i,t} \leq \theta \;\;\text{for all } i = 1, \ldots, n. ] \end{keyeq} That is: the assembly has an active capability in a type where no individual component is active.

This is a phase transition: the assembly crosses an activation threshold that no component crosses alone. It is qualitative because entirely new task categories become accessible.

\tB\; Consider an assembly of a language model (\(I_{\text{ling}} = 2.5\), \(I_{\text{spat}} = 0.05\)) and a vision encoder (\(I_{\text{spat}} = 1.2\), \(I_{\text{ling}} = 0.08\)). Neither component has \(I_{\text{eval}} > \theta\) for cross-modal evaluation tasks (interpreting diagrams with text, spatial reasoning about described scenes). But the assembly, via synergy \(K_{\text{ling,spat}} = 1.15\) and \(K_{\text{spat,eval}} = 1.2\), may achieve \(I_{A,\text{eval}} > \theta\) in cross-modal contexts---a qualitatively new capability.

\tB\; A mathematician (\(I_{\text{symb}} = 2.0\), \(I_{\text{kin}} = 0.05\)) and a sculptor (\(I_{\text{kin}} = 2.5\), \(I_{\text{symb}} = 0.08\)) may jointly produce mathematical sculptures---physical instantiations of abstract structures---that neither can produce alone. The assembly activates a capability at the intersection of symbolic and kinesthetic intelligence.

\tB\; The activation count of assembly \(A\) is [ \varepsilon_2(A) = \bigl|{t \in T : I_{A,t} > \theta \;\text{and}\; \max_i I_{i,t} \leq \theta}\bigr|. ] This counts the number of qualitatively new capabilities.

\tB\; For each type \(t\) in the \(E_2\) set, the emergent capability strength is [ I_t^{\mathrm{em}} = I_{A,t} - \theta, ] the margin above the activation threshold.

\tB\; The novelty multiplier \(\mu_t > 1\) for a newly activated type \(t\) measures the value amplification from domain access: [ \mu_t = \frac{|{\tau : R_{\tau,t} > 0}|} {|{\tau : R_{\tau,t} > 0 \;\text{and}\; R_{\tau,s} > 0 \text{ for some } s \in T_{\mathrm{active}}(A \setminus t)}|} ] where \(R_\tau\) is the requirement vector of task \(\tau\). Large \(\mu_t\) means the newly activated type opens a domain that existing capabilities cannot partially address---a genuinely new frontier.

\tB\; \begin{keyeq} [ V_2(A) = \sum_{t \in E_2(A)} I_t^{\mathrm{em}} \cdot p_t \cdot \mu_t. ] \end{keyeq}

Properties of \(V_2\)¶

\tB\; The function \(\varepsilon_2: 2^{\mathcal{E}} \to \N\) is discontinuous in the following sense: there exist assemblies \(A\) and engines \(e\) such that \(\varepsilon_2(A) = 0\) but \(\varepsilon_2(A \cup \{e\}) \geq 1\). A single component addition can trigger a phase transition.

Let \(A = \{e_1\}\) with \(I_{1,s} = 2.0\) and \(I_{1,t} = 0.05 < \theta\) for some type \(t\). Let \(e_2\) have \(I_{2,t} = 0.08 < \theta\) and \(I_{2,s} = 0.5\). If \(K_{st} > 1\) and the cross-term \((K_{st}-1) \cdot I_{1,s} \cdot I_{2,t} + (K_{ts}-1) \cdot I_{2,s} \cdot I_{1,t}\) combined with the direct contributions pushes \(I_{A \cup \{e_2\}, t}\) above \(\theta\), then type \(t\) activates. This is a discrete jump: \(\varepsilon_2\) goes from 0 to 1.

\tB\;

\(V_2\) is not submodular in general (phase transitions violate diminishing returns), but it satisfies \(\gamma\)-approximate submodularity: [ V_2(A \cup {e}) - V_2(A) \;\geq\; \gamma \bigl(V_2(B \cup {e}) - V_2(B)\bigr) \quad \text{for all } A \subseteq B, ] where \(\gamma = \theta / \max_t I_{A \cup \{e\}, t}\). This suffices for the greedy algorithm to achieve a \((1 - e^{-\gamma})\)-approximation.

[Proof sketch] Once a type \(t\) is activated (the phase transition has occurred), the marginal value of further components in type \(t\) follows the \(E_1\) pattern (additional surplus above \(\theta\)), which is submodular by Proposition ref:prop:V1-submodular. The violation of submodularity occurs only at the transition point itself, where the marginal value spikes. The parameter \(\gamma\) bounds the ratio of the spike to the post-transition marginal value. Since \(I_t^{\mathrm{em}} = I_{A,t} - \theta\) and \(I_{A,t} \leq \max_t I_{A \cup \{e\},t}\), the spike is bounded by \(\theta / \max_t I_{A \cup \{e\},t}\) times the post-transition value.

%═══════════════════════════════════════════════════════════════════════════════

Class \(E_3\): Structural Emergence¶

%═══════════════════════════════════════════════════════════════════════════════

\tC\; An assembly \(A\) exhibits structural emergence when the subobject classifier of the assembly's conceptual topos strictly extends those of its components: \begin{keyeq} [ \Omega_A \;\supsetneq\; \bigcup_{i=1}^{n} \Omega_i. ] \end{keyeq} The assembly can make logical distinctions that no component can individually make.

This connects directly to the Conceptual Irreversibility Theorem (Part V). The subobject classifier \(\Omega\) of a topos determines its internal logic---the space of "truth values" available for reasoning. A Boolean topos has \(\Omega = \{\mathbf{t}, \mathbf{f}\}\); a Heyting topos has intermediate values representing genuine vagueness. Structural emergence occurs when the assembly's truth-value space is richer than any component's.

\tC\; Consider a team of a cultural anthropologist (whose conceptual topos has a rich Heyting algebra for cultural concepts: kinship, taboo, and reciprocity carry intermediate truth values) and a formal logician (whose topos is essentially Boolean: propositions are true or false). Individually, the anthropologist cannot formalize cultural patterns (no Boolean precision), and the logician cannot perceive them (no intermediate truth values for cultural concepts). The assembly's conceptual topos may contain truth values that require both perspectives simultaneously---for instance, a formally precise statement about the degree to which a cultural practice is "taboo" that preserves the genuine vagueness of the concept while making it amenable to logical analysis. This is a new truth value: it lives in \(\Omega_A\) but not in \(\Omega_1 \cup \Omega_2\).

\tC\; The structural gap surplus of assembly \(A\) is [ \varepsilon_3(A) = \mathfrak{g}(A) - \max_i \, \mathfrak{g}(e_i) ] where \(\mathfrak{g}\) is the Heyting gap (Part V, Definition 8.1): \(\mathfrak{g} = |\Omega \setminus \mathrm{Im}(\iota)|\) counts the intermediate truth values destroyed by Booleanization.

\tC\; If the assembly's conceptual category \(\mathcal{C}_A\) strictly contains the conceptual categories of all components (i.e., \(\mathcal{C}_A \supsetneq \mathcal{C}_i\) for all \(i\)), then \(\varepsilon_3(A) \geq 0\).

This is Gap Monotonicity (Part V, Proposition 8.2) applied to the inclusion \(\mathcal{C}_i \hookrightarrow \mathcal{C}_A\). The enrichment of the concept category can only increase the number of sieves on each object, hence increase \(|\Omega|\), hence increase \(\mathfrak{g}\).

Structural emergence may be the mechanism behind paradigm shifts in the sense of Kuhn [Kuhn1962]. A paradigm shift occurs when the scientific community acquires new truth values---new ways of distinguishing "true" from "false" and "partially true"---that the prior paradigm lacked. This is precisely \(E_3\): the assembly (the community under the new paradigm) has \(\Omega_{\mathrm{new}} \supsetneq \Omega_{\mathrm{old}}\).

\tC\; [ V_3(A) = \varepsilon_3(A) \cdot p_{\mathrm{structural}} ] where \(p_{\mathrm{structural}} > 0\) is the price of logical novelty---the value of being able to think thoughts that were previously unthinkable.

Properties of \(V_3\)¶

\tC\;

The function \(V_3: 2^{\mathcal{E}} \to \R_{\geq 0}\) is not guaranteed to be submodular.

[Proof sketch] Structural emergence requires the joint topos to have truth values absent from all component topoi. This can exhibit increasing marginal returns: the third component in an assembly may unlock truth values that require all three perspectives simultaneously, producing a larger marginal gain than the second component did. This violates the diminishing-returns characterization of submodularity. A concrete counterexample: let \(e_1, e_2, e_3\) have Heyting algebras \(\Omega_1, \Omega_2, \Omega_3\) such that \(\Omega_{\{e_1, e_2\}} = \Omega_1 \cup \Omega_2\) (no structural emergence from the pair) but \(\Omega_{\{e_1, e_2, e_3\}}\) contains new elements requiring all three. Then \(\Delta V_3(e_3 \mid \{e_1, e_2\}) > \Delta V_3(e_3 \mid \{e_1\})\).

%═══════════════════════════════════════════════════════════════════════════════

Total Emergence Value and Assembly Optimization¶

%═══════════════════════════════════════════════════════════════════════════════

\tB\; The total emergence value of assembly \(A\) is \begin{keyeq} [ V(A) = V_1(A) + V_2(A) + V_3(A) = \underbrace{\sum_{t} \varepsilon_1(A,t) \cdot p_t}{ \text{quantitative surplus}} + \underbrace{\sum \cdot p_t \cdot \mu_t}} I_t^{\mathrm{em}{\text{novelty premium}} + \underbrace{\varepsilon_3(A) \cdot p. ] \end{keyeq}}}}_{\text{structural novelty}

\tA\; The marginal emergence value of adding component \(e\) to assembly \(A\) is [ \Delta V(e \mid A) = V(A \cup {e}) - V(A). ] This is the key metric for assembly optimization: it answers "how much emergence value does this component add?"

The Assembly Optimization Problem¶

\tA\; Given a budget \(B > 0\) and a catalog of available engines \(\mathcal{E} = \{e_1, \ldots, e_N\}\) with costs \(c_1, \ldots, c_N > 0\): \begin{keyeq} [ \text{maximize} \quad V(A) \qquad \text{subject to} \quad \sum_{i \in A} c_i \leq B, \quad A \subseteq \mathcal{E}. ] \end{keyeq} This is a combinatorial subset selection problem.

\tA\;

If \(V\) is submodular and monotone, the greedy algorithm---which iteratively adds the engine with highest marginal-value-to-cost ratio \(\Delta V(e \mid A) / c(e)\)---achieves a \((1 - 1/e) \approx 63.2%\) approximation to the optimal assembly.

This is the classical result of Nemhauser, Wolsey, and Fisher [NemhauserWolsey1978]. The submodularity of \(V\) ensures diminishing returns, and the greedy selection maximizes the marginal gain per unit cost at each step. The \((1-1/e)\) bound is tight for general submodular functions.

\tB\; Since \(V_1\) is submodular (Proposition ref:prop:V1-submodular) and \(V_2\) is approximately submodular (Proposition ref:prop:V2-approx-submod), the greedy algorithm provides: [nosep] - A \((1-1/e)\)-approximation for the \(V_1\) component. - A \((1 - e^{-\gamma})\)-approximation for the \(V_2\) component. - No approximation guarantee for the \(V_3\) component.

In practice, \(V_3\) is rare and typically zero for assemblies of fewer than \(\sim\)5 components, so the greedy algorithm is effective for realistic assembly sizes.

The algorithm in pseudocode: [nosep] - Initialize \(A \leftarrow \emptyset\). - While \(\exists\, e \in \mathcal{E} \setminus A\) with \(c(e) \leq B - \mathrm{cost}(A)\): [nosep] - For each candidate \(e\), compute \(\rho(e) = \Delta V(e \mid A) / c(e)\). - Add \(e^* = \arg\max_e \rho(e)\) to \(A\).

• Return \(A\).

Each iteration requires \(O(|\mathcal{E}|)\) evaluations of \(\Delta V\), each of which costs \(O(n \cdot |T|^2)\) for the synergy computation (where \(n = |A|\) and \(|T| = 8\)). Total complexity: \(O(|\mathcal{E}|^2 \cdot |T|^2) = O(64 \, |\mathcal{E}|^2)\)---negligible for realistic catalog sizes (\(|\mathcal{E}| < 100\)).

%═══════════════════════════════════════════════════════════════════════════════

The Emergence Tensor¶

%═══════════════════════════════════════════════════════════════════════════════

For a richer representation of emergence structure, we define the second-order interaction tensor.

\tB\; The emergence tensor \(\mathbf{E}(A) \in \R^{8 \times 8}\) of assembly \(A\) is defined by \begin{keyeq} [ E_{st}(A) = \frac{\partial^2 V(A)}{\partial I_{\cdot,s} \;\partial I_{\cdot,t}} ] \end{keyeq} where the partial derivatives are taken with respect to the aggregate component intelligence in types \(s\) and \(t\) (i.e., \(I_{\cdot,s} = \sum_i I_{i,s}\)).

The entries of \(\mathbf{E}\) have the following interpretation:

\begin{center}

[Table — see PDF]

\end{center}

\tB\; The emergence tensor relates to the compatibility matrix \(K\) by [ E_{st}(A) \;\approx\; (K_{st} - 1) \cdot \frac{\bigl(\sum_i I_{i,s}\bigr) \cdot \bigl(\sum_j I_{j,t}\bigr)}{V(A)} ] to first order, with additional terms from \(E_2\) (threshold discontinuities) and \(E_3\) (topos-level interactions) that \(K\) does not capture.

Differentiate \(V_1(A) = \sum_t [\sum_i I_{i,t} + \sum_{i<j,s \neq t} (K_{st}-1) I_{i,s} I_{j,t} - \max_i I_{i,t}] \cdot p_t\) twice with respect to aggregate type intensities. The direct-sum terms contribute zero to the mixed partials. The synergy terms contribute \((K_{st}-1) \cdot \partial^2 / \partial I_{\cdot,s} \partial I_{\cdot,t} [\sum_{i<j} I_{i,s} I_{j,t}]\), which yields the stated expression after normalization by \(V(A)\). The \(E_2\) and \(E_3\) contributions enter through the threshold function and the topos enrichment, respectively, neither of which is captured by the smooth \(K\)-dependent terms.

%═══════════════════════════════════════════════════════════════════════════════

Emergence Metrics Summary¶

%═══════════════════════════════════════════════════════════════════════════════

\begin{center}

[Table — see PDF]

\end{center}

%═══════════════════════════════════════════════════════════════════════════════

Three-Space Interpretation of Emergence¶

The three-space ontology (Part XIII) deepens the interpretation of all three emergence classes.

\begin{interpretation} \(E_1\) (Quantitative Emergence) as co-instantiation surplus. When two intelligence types with \(K_{st} > 1\) are co-deployed, the instantiation operator \(\Inst\) produces more \(\PS\)-structure than the sum of the individual instantiations. The surplus arises because the co-instantiation accesses \(\QS\)-modes that are entangled across both types---modes that are invisible to each type individually. The synergy factor is a measure of \(\QS\)-entanglement across intelligence types.

\(E_2\) (Qualitative Emergence) as novel instantiation. When an assembly produces capabilities not present in any member, this corresponds to the assembly's joint instantiation accessing a region of \(\QS\) that no individual member's \(\CSp\)-projection reaches. The new capability exists in \(\QS\) as potentiality; the assembly provides the composite \(\PS\)-substrate needed to project it into definite experience. Cosmological \(E_2\): the emergence of chemistry from physics, life from chemistry, and consciousness from biology are all instances of increasingly complex substrates activating previously inaccessible \(\QS\)-modes.

\(E_3\) (Structural Emergence) as ontological novelty. The strongest emergence class corresponds to instantiation producing genuine ontological novelty: \(\PS\)-structures that instantiate entirely new forms of \(\QS \to \PS\) projection. This is the emergence of new intelligence types, new filter species, or---at the cosmological scale---new spaces of possible experience. \end{interpretation}

Connections to the Framework¶

%═══════════════════════════════════════════════════════════════════════════════

The emergence theory connects to every major component of the IAG framework.

Synergy Factor and \(E_1\)¶

The synergy factor \(\mathrm{Syn}(B)\) from Part II measures the ratio of assembly capability to best-component capability. The emergence surplus \(\varepsilon_1\) measures the difference. They encode the same phenomenon at different scales: [ \mathrm{Syn}(B) = \frac{|\bI_A|}{\max_i |\bI_i|} = 1 + \frac{|\varepsilon_1(A)| + \text{(direct-sum excess)}}{\max_i |\bI_i|}. ]

Compatibility Matrix \(K\) and Prediction¶

\(K\) predicts \(E_1\) completely and provides necessary conditions for \(E_2\). Specifically: [nosep] - \(E_1\) in type \(t\) requires \(K_{st} > 1\) for some \(s \neq t\) with components strong in \(s\) and \(t\) present in the assembly. - \(E_2\) in type \(t\) requires not only favorable \(K\) entries but also sufficient aggregate sub-threshold intensity: the components' sub-threshold contributions in type \(t\) must, when amplified by cross-type synergy, cross \(\theta\). This is a necessary but not sufficient condition derivable from \(K\). - \(E_3\) requires topos-level analysis beyond \(K\).

Attention Simplex and \(E_2\)¶

The attention simplex \(\Delta^7\) (Part III) governs how the assembly allocates cognitive resources across types. \(E_2\) occurs at face activation: when the assembly's attention vector moves from a face of \(\Delta^7\) (where some types have zero allocation) to the interior (where the newly activated type receives positive attention). The hypervisor's task is to detect when \(E_2\) is imminent---when sub-threshold types are close to \(\theta\)---and shift attention to trigger the phase transition.

CIT and \(E_3\)¶

Structural emergence is the assembly-level analog of the Conceptual Irreversibility Theorem. The CIT (Part V) shows that translation between regions of the CS operator with different logics is necessarily lossy; Gap Monotonicity shows that enriching the conceptual space increases this loss. \(E_3\) is precisely the process by which an assembly enriches its own conceptual space: \(\varepsilon_3 > 0\) means the assembly has more intermediate truth values, hence a larger Heyting gap, hence greater irreversibility of translation. Structural emergence is irreversibly creative---the new distinctions, once made, cannot be unmade without information loss.

IdeaRank and \(E_2\)¶

An assembly can comprehend ideas that no component can individually comprehend. Formally: if idea \(x\) has requirement vector \(R(x) \in \\mathbb{R}^{n(e)}\) and component \(e_i\) satisfies \(I_{i,t} < R_t(x)\) for some type \(t\), but \(I_{A,t} > R_t(x)\), then the assembly comprehends \(x\) while no component does. This is \(E_2\) applied to the IdeaRank framework: the assembly's IdeaRank includes ideas that would receive zero rank from any individual component.

Hypervisor and \(V(A)\)¶

The hypervisor from Part IX manages attention across the assembly to maximize \(V(A)\). The hypervisor loss function (Definition 9.1) penalizes suboptimal attention allocation; the emergence framework gives it a concrete objective: [ \mathcal{L}{\mathrm{hyp}} = -V(A) + \lambda \cdot \mathrm{cost}(A) + \lambda_{\mathrm{time}} \cdot \mathrm{time}(A). ] Minimizing }\(\mathcal{L}_{\mathrm{hyp}}\) is equivalent to maximizing emergence value subject to cost and time constraints.

ELO and Component Contribution¶

The ELO rating system (Part IX) tracks each engine's contribution to assembly emergence over time. An engine whose addition consistently increases \(V(A)\) sees its ELO rise; one that adds cost without emergence sees its ELO fall. The ELO vector enters the scoring function as a quality-adjusted weight: [ \mathrm{score}(i, j) = \sum_t I_{i,t} \cdot R_{j,t} \cdot \sigma(E_{i,t} - \bar{E}_t) ] where \(\sigma\) is a sigmoid scaling by ELO relative to the population mean \(\bar{E}_t\).

Cognitive Temperature and Regime Transitions¶

The cognitive temperature \(T_{\mathrm{cog}}\) (Part VII) governs the transition between emergence regimes: [nosep] - At low \(T_{\mathrm{cog}}\) (focused, exploitation): the assembly operates in the \(E_1\) regime. Attention concentrates on types where the assembly is already strong. Emergence is predictable and incremental. - At high \(T_{\mathrm{cog}}\) (exploratory, exploration): the assembly explores sub-threshold types, increasing the probability of triggering \(E_2\) phase transitions. This is thermodynamically analogous to thermal activation over an energy barrier. - The optimal \(T_{\mathrm{cog}}\) balances \(E_1\) exploitation with \(E_2\) exploration---a cognitive analog of simulated annealing.

%═══════════════════════════════════════════════════════════════════════════════

Worked Example: A Three-Engine Assembly¶

%═══════════════════════════════════════════════════════════════════════════════

We illustrate the full emergence computation on a concrete assembly.

Components. Consider three AI engines with the following intelligence vectors (showing four relevant types for readability; the full 8D computation is analogous):

\begin{center}

[Table — see PDF]

\end{center}

Relevant \(K\) entries: \(K_{\text{ling,spat}} = 1.15\), \(K_{\text{spat,eval}} = 1.2\), \(K_{\text{ling,symb}} = 1.1\).

Step 1: Compute \(\bI_A\). In type eval, the individual max is \(\max(1.8, 1.5, 0.08) = 1.8\). Cross-type synergy adds terms such as \((K_{\text{spat,eval}}-1) \cdot I_{3,\text{spat}} \cdot I_{1,\text{eval}} = 0.2 \cdot 2.0 \cdot 1.8 = 0.72\) (one pair among many). After summing all pairwise cross-type contributions and the direct sum: \(I_{A,\text{eval}} \approx 1.8 + 1.5 + 0.08 + \text{synergy terms} \approx 4.5\).

Step 2: \(E_1\) surplus. \(\varepsilon_1(A, \text{eval}) = 4.5 - 1.8 = 2.7\). Similarly for other types.

Step 3: Check \(E_2\). The vision model has \(I_{3,\text{eval}} = 0.08 < \theta = 0.1\): it individually has no evaluative capability. But the assembly's \(I_{A,\text{eval}} = 4.5 \gg \theta\). Since both language models already have \(I_{\text{eval}} > \theta\), this does not count as \(E_2\) (the type was already active in \(e_1\) and \(e_2\)).

Now consider a hypothetical fourth type "cross-modal evaluation" that neither language model can perform (\(I_{1,\text{cross}} = I_{2,\text{cross}} = 0.05\)) but the vision-language assembly achieves \(I_{A,\text{cross}} = 0.3 > \theta\). Then \(\varepsilon_2 = 1\) and \(I_{\text{cross}}^{\mathrm{em}} = 0.2\).

Step 4: Total value. With \(p_{\text{eval}} = 1.0\), \(p_{\text{cross}} = 1.5\), \(\mu_{\text{cross}} = 3.0\): [ V(A) = (2.7 \cdot 1.0 + \cdots) + (0.2 \cdot 1.5 \cdot 3.0) + 0 = V_1 + 0.9 + 0. ] The \(E_2\) term (\(0.9\)) represents the value of the assembly's qualitatively new cross-modal capability.

%═══════════════════════════════════════════════════════════════════════════════

Open Problems¶

%═══════════════════════════════════════════════════════════════════════════════

[leftmargin=2em] - Higher-order synergy. The current framework uses pairwise \(K\) interactions. Three-body and higher-order emergence terms are theoretically expected but not yet formalized. Can the emergence tensor be extended to a full \(k\)-body interaction series?

• Empirical calibration of \(\mu_t\). The novelty multiplier requires knowledge of the task distribution. What is the empirical distribution of task requirement vectors in practice?

• \(V_3\) detection. Structural emergence is defined topos-theoretically but has no operational detection procedure. Can \(\varepsilon_3 > 0\) be detected from behavioral signatures of the assembly?

• Temporal emergence. Does the assembly's emergence profile change over time as components learn and adapt? How does \(V(A, t)\) evolve?

• Assembly topology. The current framework treats \(A\) as an unordered set. Does the structure of inter-component communication (who talks to whom) affect emergence? This suggests replacing \(A\) with a graph or simplicial complex.

• Negative emergence. Can assemblies exhibit \(V(A) < \max_i V(\{e_i\})\)? Under what conditions does adding components destroy value? Preliminary analysis: this occurs when \(K_{st} < 1\) terms dominate, i.e., the assembly has more interference than synergy.

%═══════════════════════════════════════════════════════════════════════════════ % References %═══════════════════════════════════════════════════════════════════════════════

References¶

See PDF for full bibliography.