Cheat-proof AI benchmarking.

Why and how

Public benchmarks leak into training data, so scores can reflect memorization instead of capability, and whoever builds the test set can cherry-pick it. Private benchmarks only move all the trust to one curator.

In peerBench, independent institutions each run a node with their own secret test data and human raters. Nodes commit hashes of their data as they create it, and later the network picks a sample of about 10% to reveal, so a node that leaks or games its data shows up as a statistical outlier. Records are public on the AT Protocol, so anyone can recompute the rankings.