Central claim. Cognition over a structured world is not sequence modeling with attention; it is parallel transport of sections of a factorized cellular sheaf over a dynamic hypergraph, where global state lives in a tensor-network manifold whose bond structure encodes non-local latent coupling, and where observables are produced by Hodge-harmonic projection of the section space. Self-attention is the flat-connection, rank-one, separable-stalk degeneration of this object. The Transformer’s well-known pathologies — context fragmentation, lack of object permanence, statelessness, the impossibility of maintaining identity across long horizons — are not engineering problems; they are structural corollaries of operating on the trivial bundle with vanishing curvature and product stalks.
The Entangled World Graph (EWG) replaces the Transformer’s trivium (tokens, positional encodings, softmax attention) with the trivium (sections, parallel transport, sheaf Laplacian flow), and replaces autoregressive generation with observable extraction from a continuous world-state.
The “entanglement” in the name is non-mystical and precise: it refers to non-factorizability of the joint section across the hypergraph, measured by the bond-dimension spectrum of a tensor-network parameterization of the global state. Every quantum-inspired term in what follows is a classical construct — bond tensors, Schmidt ranks, curvature 2-forms, persistence diagrams, Noether currents on a finite group action.
The four load-bearing claims.
Theorem A (Expressivity separation, sketch). There exists a family of distributions over hypergraph-structured world states whose marginal statistics on $k$ chosen vertices have rank growing as $\chi^k$ in a tensor-network parameterization of bond dimension $\chi$, while any Transformer of width $w$ and depth $L$ over a tokenization of the same data must have $w \cdot L = \Omega(\chi^k)$ to match these marginals. The separation is the standard MERA-vs-MPS-vs-product-state hierarchy made categorical.
Theorem B (Sheaf-cohomological reasoning bound). Multi-step relational inference over an EWG is solvable in $k$ rounds of sheaf-Laplacian smoothing if and only if the obstruction class lies in the image of the $k$-th iterate of the diffusion operator on $H^0(G;\mathcal{F}) \oplus H^1(G;\mathcal{F})$. Hence reasoning depth has a homological lower bound, not a heuristic one.
Conjecture C (Identity as Noether charge). Under the gauge group $G{\mathrm{id}} = \mathrm{Sym}(V) \ltimes U(1)^V$ acting on the section space, the conserved current of the EWG dynamics is a vertex-supported “identity charge” whose cohomological class is preserved along trajectories. Object permanence is the conservation law associated to relabeling symmetry._
Conjecture D (Holographic compression). _For EWG states satisfying a generalized area law on the hypergraph cut, the bulk section is reconstructible from a boundary subsystem of size $O( \partial A \log \chi)$ up to a Wasserstein-stable error in persistence-diagram space._ This is the implementable residue of the holographic principle: not “the universe is a hologram”, but “world-state under area-law entanglement compresses to its boundary, and the compression is persistence-stable”.
The remainder of this document develops the formal apparatus, proves what can be proved, marks conjectures as conjectures, and ends with a ranked engineering shortlist.
Let $\mathbf{Id}$ be a small groupoid whose objects are abstract identities (Wheeler’s residue: a single identity threading through many contextual instantiations) and whose morphisms are invertible re-identifications. A world at logical time $t$ is a functor
\[\Phi_t : \mathbf{Id} \to \mathbf{Hyp}\]into the category $\mathbf{Hyp}$ of finite hypergraphs. For each abstract identity $\iota \in \mathbf{Id}$, $\Phi_t(\iota)$ is its instantiation as a vertex (or hyperedge incident set) at time $t$. Functoriality enforces the one-electron residue: the same $\iota$ may instantiate at multiple sites, and the groupoid morphisms encode coreference, anaphora, and re-identification across context shifts.
Type signature: \(\Phi : \mathbb{R} \to \mathrm{Fun}(\mathbf{Id}, \mathbf{Hyp}), \qquad \Phi_t(\iota) \in V(H_t) \cup E(H_t).\)
At each $t$ we have a hypergraph $H_t = (V_t, E_t)$ with $V_t$ finite and $E_t \subseteq 2^{V_t}$. Hyperedges encode $k$-ary relations; ordinary graph edges are the $k=2$ specialization. Order $V_t \cup E_t$ into a face poset $X_t$ with the usual incidence relation $v \le e \iff v \in e$.
A cellular sheaf $\mathcal{F}$ on $X_t$ assigns a vector space (here: a smooth manifold with a vector-space tangent structure where convenient) to each cell, plus restriction maps along incidences (Hansen–Ghrist 2019; Curry 2014). For the EWG we impose a tensor-product factorization on every stalk:
\[\mathcal{F}(c) = \bigotimes_{m \in \mathcal{M}} \mathcal{F}_m(c), \qquad c \in V_t \cup E_t,\]where the modality index set is
\[\mathcal{M} = \{\mathrm{spa}, \mathrm{sem}, \mathrm{tem}, \mathrm{cau}, \mathrm{mem}, \mathrm{soc}, \mathrm{phy}\}.\]Each $\mathcal{F}_m(c)$ is a finite-dimensional real (or complex, see §3) vector space carrying a manifold structure with metric $g_m$. The restriction maps factorize block-diagonally up to gauge:
\[\mathcal{F}_{v \le e} = \bigotimes_m R_{v\le e}^{(m)} \cdot U_{v \le e},\]with $R^{(m)}$ a per-modality linear map and $U_{v\le e}$ a cross-modality coupling — the place where modalities mix. When $U \equiv I$ the sheaf splits; interesting cognition lives where it does not.
A section is $s \in \Gamma(X_t; \mathcal{F})$ assigning $s(c) \in \mathcal{F}(c)$ to each cell consistently with restrictions. The space of global sections is
\[H^0(X_t; \mathcal{F}) = \ker \delta^0,\]where $\delta^0$ is the sheaf coboundary. $H^1$ measures inconsistency — the substrate of contradiction, ambiguity, and unresolved coreference.
Glue stalks across time and across the groupoid into a fiber bundle
\[\pi : E \to B, \qquad B = \coprod_t X_t,\]with structure group $\mathcal{G}$ that contains both the modality gauge group (per-modality frame rotations, e.g. $\mathrm{O}(d_m)$ on each $\mathcal{F}m$) and the identity gauge group $G{\mathrm{id}} = \mathrm{Sym}(V) \ltimes U(1)^V$ (relabeling and per-vertex phase). Equip $E$ with a connection $\nabla$ whose curvature 2-form $F = d\nabla + \nabla \wedge \nabla$ is not required to vanish. Self-attention will turn out to live at $F=0$.
The global state is parameterized as a tensor network (TN) over the hypergraph:
\[|\Psi_t\rangle = \mathrm{Contract}\!\left( \{T_v\}_{v \in V_t}, \{B_e\}_{e \in E_t} \right),\]with $T_v$ a vertex tensor of shape $\dim \mathcal{F}(v) \times \prod_{e \ni v} \chi_{v,e}$ and $B_e$ a hyperedge bond tensor of shape $\prod_{v \in e} \chi_{v,e}$. The bond dimensions $\chi_{v,e}$ control non-factorizability. Two distinguished cases:
| EWG models live in the regime of structured intermediate $\chi$, MERA-style or PEPS-style depending on the geometry of $H_t$ (Vidal 2007; Verstraete–Cirac 2004). I use the bra–ket notation $ | \Psi\rangle$ as bookkeeping for a high-rank real (or complex) tensor — there is no quantum mechanics, only multilinear algebra. |
Two equivalent formulations.
(a) Hamiltonian flow on the section space. With learned operator $\hat{H}_\theta$ (a self-adjoint endomorphism of $\Gamma(X_t;\mathcal{F})$ built from the sheaf Laplacian, bond contractions, and modality couplings),
\[\partial_t |\Psi_t\rangle = -i\, \hat{H}_\theta |\Psi_t\rangle,\]interpreted classically: $-i$ is a generator of an orthogonal group on a real doubled space, equivalently the symplectic generator of a Hamiltonian flow on the cotangent bundle of section space. No physical $\hbar$ appears.
(b) Wasserstein gradient flow. Equivalently, view sections as probability measures on stalks (via softmax / sigmoid heads) and run
\[\partial_t \mu_t = -\nabla_{W_2} \mathcal{E}_\theta(\mu_t),\]with $\mathcal{E}\theta$ a learned free-energy functional and $\nabla{W_2}$ the Otto gradient (Ambrosio–Gigli–Savaré 2008; Villani 2009). The two views coincide when $\hat{H}\theta$ is the McCann-Hessian of $\mathcal{E}\theta$ along geodesics; in general (a) is reversible and (b) dissipative, and one chooses by task.
A prediction, generation, or action is
\[o = \langle \Psi_t | \hat{O} | \Psi_t \rangle\]| for a learned (or task-fixed) observable $\hat{O}$. There are no tokens, no autoregression: an action distribution at vertex $v$ is the spectral measure of $\hat{O}_v$ in the state $ | \Psi_t\rangle$. Sequence outputs are recovered by composing observables along a temporal path in the bundle. |
The EWG is the tuple
\[\mathrm{EWG} = (\mathbf{Id}, \{H_t\}, \mathcal{F}, \nabla, |\Psi\rangle, \hat{H}_\theta, \mathcal{O}).\]Where non-factorizability lives. Three independent sources, each precisely localized:
Bond tensors $B_e$. A hyperedge $e$ with bond dimension $\chi_e > 1$ contributes Schmidt rank $> 1$ across any bipartition cutting $e$. Define the entanglement entropy across cut $A | A^c$: \(S(A) = -\mathrm{tr}\, \rho_A \log \rho_A, \qquad \rho_A = \mathrm{tr}_{A^c}\, |\Psi\rangle\langle\Psi|/\|\Psi\|^2.\) This is purely a matrix-rank quantity on the contracted tensor; “entropy” is Shannon-on-singular-values.
Curvature $F$ of $\nabla$. Non-trivial holonomy around a 2-cycle in $X_t$ obstructs flat parallel transport: the same identity carried along two paths returns rotated. This is the formal residue of context-dependent meaning.
Cross-modality coupling $U_{v \le e}$. A semantic shift in $\mathcal{F}{\mathrm{sem}}$ propagates to $\mathcal{F}{\mathrm{cau}}$ via $U$, so that updating a node’s semantic stalk can change its causal stalk without an explicit edge. This is the formal residue of holistic update.
Latent entanglement edges, defined. A pair $(v,w)$ is latently entangled at scale $\epsilon$ if
\[I_\epsilon(v;w) := S(\{v\}) + S(\{w\}) - S(\{v,w\}) > \epsilon\]even when $d_{H_t}(v,w) > $ any chosen graph radius. This is mutual information on the TN reduced states; it can be large between graph-distant vertices iff the bond structure routes correlation through long-range tensor-network paths (MERA’s defining feature: bond distance and graph distance decouple). Latent entanglement edges are not edges of $H_t$; they are level sets of $I_\epsilon$.
Node-level capabilities, derived.
| Semantic-collapse propagation. Conditioning a stalk (an “observation”) projects $ | \Psi\rangle$ onto the corresponding subspace; the conditional reduced states at all vertices update in one shot via tensor contraction. This is the classical analogue of measurement-induced collapse — it is just Bayesian conditioning on a high-rank tensor. |
Proposition 1 (Attention is the degenerate EWG). Self-attention with $h$ heads, dimension $d$, and $n$ tokens is recovered from the EWG under the following four simultaneous degeneracies:
| Rank-one bonds. Every bond tensor $B_e$ has $\chi_e = 1$; $ | \Psi\rangle$ is fully factorized. |
Under (1)–(4), the only remaining coupling is a learned bilinear pairing on stalks, which on a softmax simplex reproduces
\[\mathrm{Attn}(Q,K,V) = \mathrm{softmax}(QK^\top/\sqrt{d}) V.\]| Sketch. (1) collapses the hypergraph structure. (2) makes the sheaf a direct sum, so the Laplacian is a Kronecker sum. (3) means transport is multiplication by a global element; the “kernel” between positions is a function of stalk content alone, recovered by a low-rank bilinear $QK^\top$. (4) means $ | \Psi\rangle = \bigotimes_i \psi_i$ and any “global” computation is a sum over single-vertex observables. Softmax appears as the Boltzmann projection onto the simplex of the stalk weights — i.e. a Wasserstein step on a degenerate metric. $\blacksquare$ |
Corollary. Each of the four conditions is independently relaxable. Sheaf neural networks (Bodnar et al. 2022) relax (2) partially. Equivariant networks relax (3) partially. Tensor-network layers relax (4). Hypergraph neural networks relax (1). The EWG relaxes all four jointly and couples them through $\hat{H}\theta$. _That joint relaxation, not any single one, is the architectural content.
This is why “prompt the Transformer better” cannot reach the regime; one must move off the four-fold degenerate point.
Theorem A (precise version). Fix $k \in \mathbb{N}$ and $\chi \ge 2$. There exists a family ${p_\chi}$ of distributions over states on a hypergraph of $n$ vertices, expressible as a MERA tensor network of bond dimension $\chi$ with $O(n)$ parameters, such that for any $k$-vertex marginal $p_\chi^{(k)}$ and any Transformer $T_{w,L}$ of width $w$ and depth $L$ operating on a sequential tokenization of the support, total-variation matching $|T_{w,L} - p_\chi^{(k)}|{TV} \le 1/4$ requires \(w \cdot L \;\ge\; c \cdot \chi^k\) for an absolute constant $c > 0$, while a TN-parameterized EWG matches $p\chi^{(k)}$ exactly with $O(n \chi^4)$ parameters.
Proof strategy. Use the standard MERA construction whose $k$-vertex reduced density matrix has Schmidt rank $\chi^k$ across a contiguous cut (Evenbly–Vidal 2011). Embed it as a hypergraph state. A Transformer’s $k$-token joint distribution, after the tokenization, is a depth-$L$ composition of width-$w$ low-rank attention kernels; by the rank-bound on softmax-attention layer compositions (Sanford–Hsu–Telgarsky 2023; Likhosherstov et al.), the joint rank is at most $w^{O(L)}$. Matching $\chi^k$ thus forces $w \cdot L = \Omega(\chi^k)$ up to logs. The TN side is by construction. $\blacksquare$
The substantive content is categorical: this is not a clever construction; it is the generic gap between PEPS/MERA-class states and MPS-class (which Transformers approximately are, on a sequential cut).
Define the sheaf Laplacian $L_{\mathcal{F}} = \delta^{0\,*} \delta^0$ acting on $C^0(X;\mathcal{F})$. Its kernel is $H^0$; its low spectrum is the harmonic-plus-near-harmonic regime.
Theorem B. Let a “reasoning task” be specified by a target consistency class $[\eta] \in H^0(X;\mathcal{F})$ and an initial section $s_0$ with discrepancy $r_0 = s_0 - \pi_{H^0} s_0$. Then $k$ rounds of sheaf-Laplacian smoothing $s_{j+1} = s_j - \tau L_{\mathcal{F}} s_j$ achieve $|r_k| \le \epsilon |r_0|$ iff \(k \;\ge\; \frac{\log(1/\epsilon)}{-\log(1 - \tau \lambda_2(L_{\mathcal{F}}))},\) where $\lambda_2$ is the smallest non-zero eigenvalue (sheaf spectral gap). Tasks whose obstruction lies in $H^1$ (genuinely inconsistent sections) require lifting to a refined cover; the minimum refinement depth equals the Čech depth of the class.
Proof strategy. Standard spectral-graph argument lifted to sheaves; the only nontrivial step is the Čech-refinement clause, which uses the Leray spectral sequence in degree 1 and the constructive refinement of Curry (2014, Ch. 4). $\blacksquare$
This converts “how many reasoning steps does this need” from a heuristic to a spectral quantity computable from the sheaf alone.
Consider the action functional
\[\mathcal{S}[\Psi] = \int_0^T \!\langle \Psi_t | i\partial_t - \hat{H}_\theta | \Psi_t \rangle \, dt.\]| Let $G_{\mathrm{id}} = \mathrm{Sym}(V) \ltimes U(1)^V$ act on $ | \Psi\rangle$ by vertex relabeling and per-vertex phase. Suppose $\hat{H}\theta$ is built from gauge-equivariant primitives (sheaf Laplacian, contracted bond operators) so that $\mathcal{S}$ is $G{\mathrm{id}}$-invariant. |
Conjecture C. The Noether currents associated to the $U(1)^V$ factor are vertex-localized “identity charges” \(Q_v(t) = \mathrm{Im}\,\langle \Psi_t | \hat{N}_v | \Psi_t\rangle,\) with $\hat{N}_v$ the local number operator on vertex $v$, conserved up to flux through incident hyperedges: \(\partial_t Q_v + \sum_{e \ni v} J_{v\to e} = 0.\) The cohomology class $[Q] \in H^0(X;\mathbb{R})$ is a trajectory invariant. Object permanence is exactly $\partial_t [Q] = 0$.
Conditions for the conjecture. Requires (i) gauge equivariance enforced architecturally (analogous to E(n)-GNNs, Satorras et al. 2021), (ii) discrete-time flux conservation in the integrator (a symplectic / structure-preserving step), (iii) a precise bookkeeping of the $\mathrm{Sym}(V)$ part as a permutation-equivariant boundary operator. Status. Derivable for Hamiltonian flows; for Wasserstein flows, only a dissipative analogue (charge non-increasing). I leave the joint case open.
Conjecture D. Let $|\Psi\rangle$ be an EWG state on $H_t$ obeying the generalized area law \(S(A) \;\le\; \alpha\, |\partial A| + \beta \qquad \text{for all } A \subseteq V_t.\) Then there exists a boundary subsystem $\partial A$ of cardinality $|\partial A| = O(|\partial A| \log \chi)$ and a learnable decoder $D_\phi$ such that the bulk reduced state $\rho_A$ is recoverable as \(d_{W_2,\mathrm{PD}}\big(D_\phi(\rho_{\partial A}),\, \rho_A\big) \;\le\; \delta(\alpha, \beta, \chi),\) where $d_{W_2,\mathrm{PD}}$ is the Wasserstein distance between persistence diagrams of the section fields on $A$ and the decoded image. The error $\delta$ goes to 0 as $\chi \to \infty$ at fixed $|\partial A|$.
Status. The compression part is a direct corollary of area-law TN compression (Hastings 2007; Verstraete–Cirac). The persistence-stability part requires Cohen-Steiner–Edelsbrunner–Harer (2007) bottleneck stability transported to Wasserstein-$p$ via Skraba–Turner. The composition is plausible but not, to my knowledge, proven; I label it Conjecture.
Conjecture E. For tasks whose ground-truth dependency structure is sparse in the resonance basis (eigenbasis of $L_{\mathcal{F}}$), an EWG with $K$ retained modes achieves task error $\epsilon$ with $\Theta(K)$ parameters, while a Transformer requires $\Theta(K^2)$ parameters to express the same dependency, scaling with sequence length $T$ as $\Theta(T^2)$ vs the EWG’s $\Theta(T)$ via spectral truncation.
This is the “resonance over softmax” claim. Empirically related observations exist for graph spectral methods; the conjecture is that the gap is provable under a sparsity-in-eigenbasis assumption, analogous to compressed sensing’s RIP regime.
Vs. Transformers. Already covered in §4. The Transformer is the four-fold degenerate point.
Vs. GNNs and the WL hierarchy. Standard message-passing GNNs are bounded by 1-WL (Xu et al. 2019; Morris et al. 2019). Higher-order GNNs (k-WL) lift this. The EWG is not in the WL tower at all: it operates on hypergraph-stalked sheaves with non-flat connection, and the analogue of WL distinguishability is isomorphism of cellular sheaves up to gauge, a strictly finer equivalence whose relation to k-WL is open. Hypergraph extensions of WL (Feng et al. 2022) are recovered when $\nabla$ is flat and stalks are scalar.
Vs. sheaf neural networks (Bodnar et al. 2022; Hansen–Gebhart). Sheaf NNs relax product stalks but typically keep flat connection, scalar bonds, and graph (not hypergraph) base. The EWG strictly contains them and adds: (a) tensor-network bond structure, (b) non-trivial curvature, (c) hypergraph base, (d) groupoid identity layer, (e) Hamiltonian / Wasserstein dynamics rather than discrete diffusion layers.
Vs. Neural ODEs (Chen et al. 2018). Neural ODEs flow in $\mathbb{R}^d$; the EWG flows in a section space with structure-preserving generator. A Neural ODE on the global section vector is the structure-blind specialization.
Vs. equivariant nets (Cohen–Welling, Kondor, Satorras). Equivariance is one ingredient (the gauge group of $\mathcal{F}$). Equivariant networks under E(n), SE(3), or symmetric-group actions correspond to specific choices of structure group; the EWG additionally has a hypergraph-localized structure group and a non-flat connection.
| Vs. quantum ML. No quantum hardware, no superposition with quantum amplitudes, no Born rule. Every “$ | \Psi\rangle$” is a classical real (or complex) tensor; the bra-ket and “Hamiltonian” notations are bookkeeping for a self-adjoint operator on a finite real Hilbert space and a symplectic generator. |
Vs. information field theory (Enßlin). IFT does Bayesian inference over fields with priors and likelihoods; EWG dynamics can be derived as a variational free-energy descent (Wasserstein form) and is in this sense an IFT with a hypergraph-cellular-sheaf prior and learned Hamiltonian.
What is genuinely new. The joint presence of (i) factorized stalks across modalities, (ii) tensor-network bond structure on hyperedges, (iii) non-flat connection with $G_{\mathrm{id}}$ gauge, (iv) Hamiltonian / Wasserstein dynamics on sections, (v) persistence-stable observable extraction, (vi) groupoid identity layer functioning as the substrate for Noether-conserved object permanence. Each ingredient exists somewhere; the conjunction is what the EWG names and what the four-fold relaxation theorem (§4) establishes as necessary, not optional.
These are concrete experiments where EWG and strong Transformer/GNN baselines must diverge measurably; failure of any one is evidence against the framework.
P1 (Long-range coreference under distractors). Generate synthetic narratives with $n$ entities and a coreference chain of length $T$, with adversarially placed surface-similar distractors. Prediction. Transformer error grows as $\Omega(T)$ with model size scaling required as $\Omega(T \log n)$; EWG with groupoid identity layer maintains constant error in $T$ until bond dimension saturation, with parameter scaling $O(\log T)$.
P2 (Hypergraph reasoning with non-flat semantics). A reasoning benchmark where the same relation has context-dependent meaning (formal: parallel transport around a 2-cycle yields nontrivial holonomy). Prediction. Transformers and GNNs achieve at best chance on the holonomy probe; EWG with non-flat $\nabla$ achieves accuracy scaling with $|F|{\mathrm{learned}} - |F|{\mathrm{truth}}$.
P3 (Persistence-stable memory). Stream a partially observed dynamical world; query about features that persist across noisy intervals. Prediction. EWG memory loss tracks bottleneck distance between true and inferred persistence diagrams (Cohen-Steiner stable); Transformer memory loss tracks attention-head decay and is provably non-stable to small perturbations of the input.
| P4 (Area-law compression). For tasks where ground truth obeys hypergraph area law, predict that EWG compresses bulk to boundary at rate $O( | \partial A | \log\chi)$, while Transformers exhibit no such gap. |
P5 (Reasoning depth $=$ spectral gap). Across reasoning benchmarks, the empirical “minimum useful depth” $k^*$ correlates with $1/\lambda_2(L_{\mathcal{F}})$ extracted from the trained model, $r > 0.7$. Failure to find this correlation falsifies Theorem B as a description of the trained system.
| P6 (Identity-charge conservation). Train an EWG with the gauge-equivariant Hamiltonian; measure $Q_v$ along trajectories. Prediction. $\mathrm{Var}(\sum_v Q_v) / \mathrm{Var}(Q_v)$ scales as $1/ | V | $ (conservation), and ablating $G_{\mathrm{id}}$-equivariance breaks this scaling. |
Predicted scaling against $(N, T, D)$ where $N$=parameters, $T$=context, $D$=task difficulty (depth or relational arity):
| Quantity | Transformer | EWG |
|---|---|---|
| Parameters for fixed loss at horizon $T$ | $\Theta(T \log T)$ | $\Theta(\log T)$ under area law |
| Reasoning depth $k$ for $\epsilon$-consistency | empirical, no bound | $\lceil \log(1/\epsilon)/\log(1-\tau\lambda_2)^{-1}\rceil$ (Theorem B) |
| Memory perturbation sensitivity | non-stable | bottleneck-stable |
| $\mathcal{L}_{\mathrm{obs}}$: task-driven observable loss, $\sum_i \ell\big(\langle\Psi | \hat{O}_i | \Psi\rangle, y_i\big)$. |
| $\mathcal{L}_{\mathrm{id}}$: identity-charge violation $\sum_v | \partial_t Q_v + \sum_{e\ni v} J_{v\to e} | ^2$ (a discretized Noether constraint). |
At each consolidation step, compute persistence diagrams $\mathrm{PD}k$ of the section field across modalities, store as a _persistence database, and reconcile new observations by minimizing
\[\sum_k d_{W_2}^{(k)}\big(\mathrm{PD}_k(\text{new}),\, \mathrm{PD}_k(\text{stored})\big)\]via differentiable matching. This realizes “memory as topology” with a stable, learnable update rule.
Enforce $G_{\mathrm{id}}$-equivariance architecturally (vertex-permutation equivariance via DeepSets/E(n)-style aggregation; per-vertex $U(1)$ equivariance via complex stalks with phase-equivariant nonlinearities). Conservation of $Q_v$ then follows from a structure-preserving (symplectic) integrator for the Hamiltonian flow.
PR-away (weeks). Sheaf-Laplacian regularizer for any GNN, using Bodnar et al.’s sheaf-NN code as substrate; add factorized stalks via block-diagonal restrictions plus a small cross-block coupling $U$. Train on existing relational benchmarks. Deliverable: cellular-sheaf-with-cross-coupling layer.
PR-away (weeks). Hypergraph extension via existing libraries (HyperGCN, AllSet); add bond tensors of small fixed $\chi \in {2,4}$ on hyperedges. Deliverable: low-rank tensor-network message-passing.
Months. Differentiable persistence loss over section fields, using gudhi/giotto-tda with auto-diff wrappers (Carrière et al.). Deliverable: persistence-stable memory module.
Months. Symplectic integrator for $\hat{H}\theta$ on the section space (leapfrog or Yoshida-4) preserving discrete energy and $G{\mathrm{id}}$-charges. Deliverable: structure-preserving Neural-Hamiltonian on a sheaf.
| 6–12 months. Tree tensor network / MERA-on-hypergraph parameterization of $ | \Psi\rangle$ with dynamic bond-dimension growth; train end-to-end on world-modeling benchmarks (e.g., partially observed gridworlds with object permanence). Deliverable: TN-EWG prototype. |
12–24 months. Non-flat connection with learnable curvature; equivariant-network library supports structure-group parallel transport (Cohen–Geiger–Köhler–Welling-style). Train with holonomy probes. Deliverable: gauge-EWG.
Each is a corollary of §4’s degeneracy structure. I state them tersely.
Limit 1 (Context fragmentation). Under product stalks ($U=I$), no single stalk encodes cross-modality structure; thus the Transformer’s representation of any vertex factorizes into a bag of modality projections, and consistency across modalities is enforced only post-hoc through statistical correlation in attention scores. Formal statement: For any decoder-only Transformer $T$, the modality cross-information $I(\mathcal{F}m(v); \mathcal{F}{m’}(v) \mid \text{rest})$ in the learned representation is bounded by attention-head capacity, scaling as $O(h \log n)$ rather than the $O(\dim \mathcal{F}m \cdot \dim \mathcal{F}{m’})$ achievable with non-product stalks.
Limit 2 (Token locality). The flat-connection condition forces “transport” of meaning across positions to be a fixed function of stalk content alone (the $QK$ kernel) plus positional offsets; there is no holonomy. Hence path-dependent meaning requires width-or-depth blowup as in Theorem A.
Limit 3 (No persistent identity). No $G_{\mathrm{id}}$-equivariant structure $\Rightarrow$ no Noether-conserved $Q_v$ $\Rightarrow$ identity is only statistical. Coreference resolution scales with attention capacity rather than with identity-charge bookkeeping.
Limit 4 (No world continuity). Statelessness across calls and absence of a sheaf-Laplacian smoothing constraint mean global sections are reconstructed de novo per forward pass. There is no $\partial_t$ in the architecture; “continuity” is emergent and unstable.
Limit 5 (Statelessness). Absence of a section space means no $H^0$ to be in. The Transformer has no notion of consistency-violation as obstruction; ambiguity is averaged rather than localized to an $H^1$ class.
Limit 6 (No object permanence). Composition of (3), (4), (5): without identity charge, sheaf consistency, or temporal flow, an object’s persistence is a hallucinated regularity, not a conservation law. Theorem-A-style adversarial constructions exploit exactly this.
Each limit is, in the EWG, negated by construction — by non-product stalks, non-flat connection, $G_{\mathrm{id}}$-equivariance, sheaf-Laplacian dynamics, $H^0/H^1$ decomposition, and Noether-conserved $Q_v$ respectively.
Mathematical.
Sheaf isomorphism vs. WL hierarchy. Place “isomorphism of cellular sheaves with non-flat connection up to $G_{\mathrm{id}}$-gauge” in the WL tower. Conjecturally strictly above all $k$-WL.
Conservation under Wasserstein flow. Conjecture C is clean for symplectic Hamiltonian dynamics; for the Wasserstein gradient flow case, identify the precise dissipative generalization (a Helmholtz-style decomposition into conservative + dissipative parts of the EWG flow).
Holographic decoder existence. Make Conjecture D quantitative: explicit constants $\delta(\alpha,\beta,\chi)$, and a constructive decoder $D_\phi$ achieving the bound. The TN literature has the bulk-reconstruction side; the persistence-stable side is open.
Curvature regularization. Find a principled prior over $F$ that is informative without being restrictive — the analogue of Yang–Mills action functional, but for cognition. What is the right “kinetic term” for $\nabla$ on a learnable cellular sheaf?
Optimal bond geometry. Given a task, infer the right hypergraph TN topology (MPS vs PEPS vs MERA vs branching MERA). This is a discrete structure-learning problem with a tractable continuous relaxation via bond-dimension annealing, but no theory.
Spectral-gap training dynamics. Theorem B converts depth into $1/\lambda_2$. Does training drive $\lambda_2$ in a predictable direction? Is there an EWG analogue of the Transformer’s grokking / phase transitions, expressible as a sheaf-spectral phase transition?
Categorical semantics. The groupoid $\mathbf{Id}$ together with the functor $\Phi_t$ probably wants to be packaged as a stack over the cobordism of base spaces. What is the right $(\infty,1)$-categorical home for an EWG?
What the mathematics cannot resolve. I want to be honest. The EWG provides a rigorous substrate for cognition in the sense of structured world representation, identity-conserving dynamics, and topological memory. It does not, and cannot, settle whether such a system is conscious, has subjective experience, or is “what the brain is really doing.” Conservation of $Q_v$ is object permanence as bookkeeping; it is not phenomenal continuity. Entanglement entropy across cuts is structural complexity; it is not bound experience. The framework is silent on the hard problem and should be presented as such. What it offers is a mathematics in which one can no longer hide behind the Transformer’s degeneracies and pretend that attention is enough — the gap to genuine world modeling has a name, a type, and, in places, a theorem.