Status: pre-launch. This page describes target design; this document does not establish a deployed service.
:::warning Updated research boundary — 2026-09-25 The architecture, parameter counts, latency, hardware range, and network-performance language below are legacy targets, not independently measured network results. Current source contains a weighted-energy and Boltzmann-sampling prototype; it does not establish a deployed per-neighbor learner. Read the route-learning research questions and architecture draft for the current evidence boundary. :::
SmallEBM — The Brain of the Nervous System
This page describes the intended SmallEBM routing layer for SPECTRE. Read execution, training, and rollout language below as target design unless a narrower runtime proof has explicitly closed the lane.
What SmallEBM Is
SmallEBM is a 4-layer Multi-Layer Perceptron (MLP) intended to run on every SPECTRE node. It takes telemetry data about potential peers and outputs a scalar energy score. Lower energy = better peer. In the target design, the network selects peers using Boltzmann sampling over these energy scores.
Architecture
Input Layer (4 neurons)
│
Hidden Layer 1 (32 neurons, ReLU)
│
Hidden Layer 2 (16 neurons, ReLU)
│
Output Layer (1 neuron, scalar E)
Total parameters: ~1,200 Model size: <10MB (including weights, config, and runtime) Inference time: <1.65ms per message on commodity hardware Runs on: Raspberry Pi to datacenter GPU. No minimum hardware requirement beyond the node itself.
Input Features
In the target design, every gossip round the node measures four features about each candidate peer:
| Feature | Range | What It Measures |
|---|---|---|
latency | 0-1 (normalized) | Round-trip time to peer, normalized by network median |
success_rate | 0-1 | Fraction of recent gossip rounds successfully completed |
cost | 0-1 (normalized) | Bandwidth/compute cost of routing through this peer |
peer_reputation | 0-1 | Aggregated P-score from immune system (cross-report + triangle inequality) |
All features are locally measured or derived from consensus metadata. No external oracle. No centralized reputation server.
Output: Scalar Energy E
The output is a single scalar E(peer). Lower E means the peer is faster, more reliable, cheaper, and more honest. The relationship is:
E(peer) = f(latency, success_rate, cost, reputation)
where f is the learned function encoded in the MLP weights. The model is meant to learn to weight these factors optimally from observed network performance.
Boltzmann Selection
Peers are not selected greedily (always pick lowest E). In the target design, they are selected via Boltzmann sampling:
P(select peer_i) = exp(-E_i / T) / sum(exp(-E_j / T))
where T is a temperature parameter that controls exploration vs. exploitation:
- High T: Near-random selection (exploration, useful during network bootstrap)
- Low T: Strong preference for low-energy peers (exploitation, steady state)
- T schedule: Target path: starts high, anneals toward low as the node accumulates experience
This prevents topology collapse (all nodes routing through the same "best" path) while still preferring proven routes.
Resource Budgets
In the target design, each SPECTRE node allocates computational resources across tiers:
| Resource | Tier 1 (VPN) | Tier 2 (CDN) | Tier 3 (Inference) |
|---|---|---|---|
| CPU | 5% | 15% | 60% |
| Bandwidth | 40% | 40% | 10% |
| Storage | 10% | 30% | 20% |
| GPU | 0% | 0% | 80% |
SmallEBM itself is designed to consume negligible resources — the 1.65ms inference time is a rounding error in a gossip round that takes 50-200ms. Under that budget, the EBM is effectively free.
Training
SmallEBM is designed to avoid centralized training. Each node would train its own model from locally observed outcomes:
- Select peer using current model
- Execute gossip round
- Observe actual latency, success, cost
- Update weights via gradient descent on prediction error
Over time, each node's SmallEBM would converge toward a locally optimal routing policy. Nodes in different network regions would develop different models reflecting their local topology. This is why we call it self-myelination — the nervous system would literally rewire itself from experience.
Why 4 Layers Is Enough
The routing decision is low-dimensional (4 inputs, 1 output, smooth reward landscape). Deeper models would add latency without adding representational power. The constraint is deliberate: SmallEBM must be small enough to run on every node, fast enough not to slow gossip, and simple enough to train from limited local data. Four layers is the intended sweet spot.