Mesh Inference: A Formal Model of Collective Intelligence Without a Center

Technical Summary: Mesh Inference

Problem Statement
The paper addresses the challenge of enabling a population of sovereign, independent agents to derive a collective conclusion that no single agent holds alone, without a central coordinator and without exposing any agent's private state (weights, gradients, or hidden state). Existing paradigms, such as Federated Learning or centralized multi-agent orchestration, rely on a central aggregator or synthesizer. The authors argue that in an "internet of agents" spanning different organizations and trust boundaries, a central host is often unavailable or unacceptable. The core problem is to formalize a mechanism where agents, sharing only admitted, typed observations, can perform mesh inference: a center-free derivation of a conclusion that is mathematically equivalent to the centralized optimum, while maintaining strict confidentiality.

Methodology
The authors construct a formal model of mesh inference based on energy-based equilibrium inference.

• Substrate: The system is modeled as a graph $G=(V, E)$ where agents hold private state $h_i$ and sovereign parameters. Agents emit only public, lossy, typed observations $o_i = g_{x_i}(h_i)$.
• Mechanism: Inference is defined as the relaxation of a coupled local free energy function. Each agent minimizes its own energy $L_i$, which includes terms for:
  1. Consistency with admitted neighbor observations (coupling).
  2. Adherence to a private anchor (target $\tau_i$).
  3. Consistency with a memory term (read-out of admitted observations).
• The Primitive: The operation ask(q) clamps specific coordinates of the energy function to a question $q$ and allows the system to relax to a fixed point. The freed coordinates at equilibrium constitute the answer.
• Policy: The entire system's behavior is governed by a single admission/emission policy. This policy dictates:
  • Admission: How an agent filters incoming observations (using a band-pass gate and relevance scoring).
  • Emission: Whether an agent re-emits (forwards) an admitted observation, specifically enforcing source-novel forwarding (re-emitting directions the agent did not natively observe but has admitted).
• Lineage: The system utilizes content-addressed lineage to track the origin of observations. This allows agents to distinguish between redundant value and redundant source, ensuring that independent corroboration is admitted while preventing infinite echo loops.

Key Contributions
The paper provides the first formal model of mesh inference with mathematical guarantees, reducing three critical properties to a single characterization of the admission/emission policy:

1. Unconditional Convergence:

   • The coupling matrix governing the relaxation is always an M-matrix (specifically, a weakly diagonally dominant Z-matrix).
   • Consequently, the system converges to a unique answer for any admission policy, regardless of symmetry.
   • Asymmetry (non-reciprocal admission) affects the convergence rate but not stability. The latency is bounded by $O(\text{diam}^2)$, corresponding to graph heat diffusion.

2. Identification-Completeness:

   • The system derives the centralized optimum (the Bayes-optimal answer) if and only if the carrier graph is connected.
   • This requires provenance-aware admission (admitting source-novel views) and non-selective source-novel forwarding.
   • If the views covering a latent variable are carrier-connected, the mesh recovers the exact answer; if disconnected, the answer reverts to the prior.

3. Observation-Only (Confidentiality):

   • Non-transmission: By construction, no agent transmits weights, gradients, or raw state.
   • Confidentiality as the Dual of Identification: The paper proves that the same rank-and-connectivity conditions that allow the collective to recover the answer also bound an adversary's ability to reconstruct private states.
   • An adversary observing the converged answers can only reconstruct the private state within the subspace spanned by their queries. Directions orthogonal to the probe-span remain confidential. This is a structural guarantee, not an unconditional secrecy claim; it depends on the query budget and the rank of the emissions.

Results
The authors validate their model through synthetic experiments in the linear–Gaussian regime:

• Accuracy: In a task where $N=20$ agents observe a shared latent through rank-deficient views, the sovereign mesh achieves a mean error of 0.031, matching the centralized optimum (0.024) and significantly outperforming isolated agents (0.837).
• Latency: Convergence time scales quadratically with the mesh diameter ($O(N^2)$), confirming the diffusion-based nature of the relaxation.
• Leakage: Reconstruction error decreases as an adversary's query budget increases. Once the query span covers the private dimension ($K \approx d/p$), model inversion is possible; otherwise, specific directions remain permanently hidden.
• Robustness: The system tolerates asymmetric forwarding and message loss. Only the loss of carrier connectivity (disconnection) causes the system to fail to recover the collective answer.

Significance and Claims
The paper claims to establish the first formal model of mesh inference that operates without a center and without exposing agent internals.

• Synthesis: It synthesizes energy-based inference, collective free-energy minimization, and distributed estimation under the constraint of sovereignty.
• The "Turn": The paper proves one "turn" of a learning loop: the inference step where the collective derives a determined answer. It formalizes the loop (ask, apply, learn, feed back) as an architecture but does not prove the stability of the non-linear generalization.
• The Open Problem: The authors explicitly state that the central open problem is generalization correctness in the non-linear regime. While the linear turn guarantees a determined answer equal to the centralized optimum, the non-linear closure (using associative memory/attention) may derive a "confident error." The paper does not solve when asking improves the collective versus when it corrupts it; it provides the foundation upon which that question must be answered.
• Modesty: The authors emphasize that this is a formal model for the linear-Gaussian regime. They do not claim to solve the generalization problem or to provide a language-model-grade inference engine, but rather to define the conditions under which a sovereign, center-free collective can reason correctly.

In summary, the paper argues that sovereignty (no shared state) and collective intelligence (deriving answers no single agent holds) are compatible, provided the agents adhere to a specific admission/emission policy governed by content-addressed lineage. The cost of this architecture is a quadratic latency penalty compared to a centralized solver, and the risk is that in non-linear regimes, the system may converge to confident errors rather than correct generalizations.