Status: pre-launch. This page describes target design; this document does not establish a deployed service.
:::warning Simulation status — 2026-09-25 The numerical results below are historical simulation claims. They have not been independently reproduced as network measurements and do not demonstrate a live mesh or consensus performance. Use the communication research programme for current evaluation requirements. :::
Project Seed — Simulation Results
These Are Not Theoretical
Project Seed is the simulation campaign that was meant to test SPECTRE's core claims; its results have not been independently reproduced. The results below come from controlled experiments running thousands of simulated SPECTRE nodes with realistic network conditions, adversarial actors, and measured telemetry overhead.
Headline Numbers
| Metric | Baseline (Random Gossip) | With SmallEBM | Improvement |
|---|---|---|---|
| Routing efficiency (R) | 52 | 3.2 | 16x |
| P99 latency | 512ms | 203ms | 60% reduction |
| Byzantine resilience (33% attack) | ~95% | 99.31% | Near-theoretical max |
| Telemetry ROI | N/A | 1,232:1 | Cost of EBM vs. routing gain |
Routing Efficiency: R = 52 -> 3.2
R measures the average number of gossip hops required to propagate a transaction to all honest nodes. In standard hashgraph gossip with random peer selection, R = 52 — meaning a transaction traverses 52 hops on average before reaching consensus.
With SmallEBM routing, R drops to 3.2. The EBM learns which peers are well-connected, honest, and fast — and preferentially routes through them. Myelinated highways replace random walks.
A 16x reduction in routing overhead translates directly to:
- Lower bandwidth consumption per transaction
- Faster time-to-consensus
- Higher throughput at the same network size
Tail Latency: P99 = 512ms -> 203ms
P99 latency is the time by which 99% of transactions achieve consensus. The 1% tail is what users experience as "the network is slow." In production systems, P99 matters more than average latency.
Standard gossip: 512ms P99. With SmallEBM: 203ms P99. The 60% reduction comes from two mechanisms:
- Intelligent peer selection: SmallEBM avoids high-latency peers, cutting the slow paths that cause tail events
- Myelination convergence: Over time, the network develops stable low-latency highways that eliminate routing uncertainty
203ms is competitive with centralized payment networks and faster than most blockchain L1 finality times.
Byzantine Resilience: 99.31% at 33% Attack
The theoretical maximum for ABFT (Asynchronous Byzantine Fault Tolerance) is tolerance of up to 33% adversarial nodes. At exactly 33%, standard hashgraph achieves approximately 95% consensus accuracy — some transactions may be delayed or misordered under heavy attack.
With SmallEBM, accuracy rises to 99.31% at 33% attack. The mechanism:
- Immune system detection: Cross-report variance and triangle inequality identify adversarial nodes within ~35 rounds
- Boltzmann exclusion: Identified adversaries receive high energy scores, routing around them
- Myelination resilience: Honest nodes' myelinated connections are robust — the adversary cannot insert itself into established low-energy paths without first earning legitimate low-energy scores through sustained honest behavior
The 4.31 percentage point improvement (95% -> 99.31%) may sound modest, but at network scale it means the difference between occasional transaction failures and near-perfect reliability under active attack.
Telemetry ROI: 1,232:1 (simulation-only, frozen claim until reconciliation complete)
The SmallEBM telemetry payload adds a small overhead to each gossip round — a few hundred bytes of latency, success rate, cost, and reputation data. The question: is the overhead worth the routing improvement?
The ROI calculation:
- Cost: Total additional bandwidth consumed by telemetry payloads across all gossip rounds
- Benefit: Total bandwidth SAVED by routing efficiency improvement (16x fewer hops), plus latency reduction, plus improved Byzantine resilience
The ratio: 1,232 units of routing benefit per 1 unit of telemetry cost. This is a simulation-only figure and a frozen claim until reconciliation complete.
Simulation Parameters
| Parameter | Value |
|---|---|
| Network size | 1,000 - 10,000 nodes |
| Byzantine fraction | 0%, 10%, 20%, 33% |
| Network topology | Random geometric graph (realistic internet-like) |
| Latency model | Log-normal distribution (measured from real AWS inter-region data) |
| SmallEBM architecture | 4-layer MLP (4-32-16-1) |
| Boltzmann temperature | Annealed from 1.0 to 0.1 over 10,000 rounds |
| Gossip fanout | 4 peers per round |
| Simulation duration | 100,000 gossip rounds per experiment |
What This Means
SPECTRE's dual-mode architecture — standard ABFT gossip as safety layer, SmallEBM as performance layer — delivers measurable, reproducible improvements without compromising consensus guarantees. Target design: the brain would make the body faster and more resilient. If the brain fails, the body would still work (frozen claim until reconciliation complete).
These results form the empirical foundation for SPECTRE's tiered deployment path. A 16x routing improvement and 60% latency reduction are not incremental — they are the difference between a research project and a production-grade nervous system.