Peer-to-Peer
What peer-to-peer means here
AETRON runs as a peer-to-peer network of Miners. Every Miner is an equal participant that both does work and helps check the work of others. There is no central server that requests must pass through, and no separate validator role that grades Miners from above. The Miners themselves form the network.
How requests reach Miners
Each Neuronet has its own discovery and gateway routing, configured by the Owner. For self-hosted setups this includes a gateway address. When a request comes in, the discovery and gateway layer routes it to the Miners running that Neuronet, and one of them serves the response. Adding or removing Miners does not require slot assignments or registration limits, so a Neuronet can grow or shrink with demand.
How Miners check each other
A Miner does two things at the same time. It serves requests (execution), and it re-checks a portion of other Miners' work (verification). The protocol decides who re-checks whom and enforces it through the reward rules, so verification is not optional and it is not handed to a privileged class of node. Both jobs feed into what a Miner earns. Honest work is confirmed by peers, and results that do not hold up under re-checking do not get paid.
Correctness is measured by Proof of Intelligence, and throughput is measured by Pulse, a proof-of-compute signal that works the same across NVIDIA, AMD, Apple Silicon, and CPU. Capacity follows the stake behind a Neuronet rather than a dialed setting, and work beyond that staked capacity pays nothing.
Joining the mesh
The transport is libp2p: Noise for the handshake, Kademlia for discovery, and circuit relay for miners that sit behind NAT. On the reference client the listener is off until mesh_listen names an address such as /ip4/0.0.0.0/tcp/4101; without it the miner has no way to receive a gateway's job and only sees what the chain tells it.
Bootstrap and relay addresses are built into the client per network, so a normal miner sets nothing. Each network has its own relays, and each is published under both a DNS and a raw IP multiaddr so that a relay can move hosts without breaking clients that cannot resolve DNS. Hijacking one of those addresses gains nothing, because the peer id is checked during the Noise handshake; the worst case is a refused connection.
The Kademlia protocol name carries the network id, which keeps mainnet and testnet peers in separate namespaces. Overriding AETRON_NETWORK_ID by hand puts you in a namespace of your own where nobody else is listening. See Configuration for the fields and environment variables involved, and Troubleshooting for the discovery failures that come up in practice.
Why the design is resilient
Because every Miner can serve and verify, no single node is in control and no single node is required. If some Miners go offline, the rest keep serving requests and keep checking each other. The reference client (aetron-miner) connects to a node, registers, serves inference, and submits proof artifacts, but the protocol is open: anyone running a compatible client per the public specs is a full participant, not a second-class one.