Peer Validation

In a network of strangers’ GPUs, one question matters more than any other: is this node actually running the model it says it is? A node could claim to serve a 27B model and quietly answer with a tiny one to save money. Peer Validation is how we keep everyone honest.

Trust is earned, not assumed

When you register an instance, the network doesn’t take your word for it. Your node starts at baseline payout and earns its way up. Reputation tiers gate access to the highest emission multipliers, so the rewards for sticking around and serving real tokens compound over time.

How we check

Validation happens on two timescales:

Online (live). Lightweight, continuous checks woven into normal traffic — statistical analysis of output distributions (logprobs) and occasional semantic-consensus replicas, where the same prompt is quietly served by more than one node and the answers are compared. A node pretending to run a model it isn’t will drift from the expected fingerprint.
Offline (audits). Asynchronous Proof of Useful Work audits run in the background to confirm a node’s claimed model and capacity over time.

Why cheating doesn’t pay

The math is deliberately stacked toward honesty:

A spoofing node never accumulates the validated history needed to reach higher trust tiers, so it’s locked out of the biggest bounties.
Getting caught means leaving rotation. The effort of faking it earns less than simply serving real requests would have.
Long-term honest participation is exponentially more profitable than short-term tricks.

What operators should expect

Your first validated tokens matter most — they’re how trust starts accruing. Keep your node healthy and online.
You don’t need to do anything special for validation; it runs alongside normal traffic.
A hardware or model upgrade is reflected automatically as the network observes your new behavior.

For the full economic context around trust tiers and payouts, see The Fair Ecosystem.