We measured trust on ERC-8004 — and the naive average crowns the wrong agent

Short summary

ERC-8004 provides on-chain identity and reputation registries for AI agents but leaves scoring to applications. The authors show that naive average scoring fails on real Base mainnet data — it rewards volume of correlated evidence over breadth of independent evidence. They implement Jøsang & Ismail's Beta Reputation System with subjective-logic operators, producing expectation and uncertainty as first-class outputs, fixing the correlated-evidence problem with a two-counter model.

• •Naive average scoring on ERC-8004 crowns the wrong agent by rewarding correlated evidence
• •Beta Reputation System (2002) models trust as a Beta posterior with uncertainty as a first-class output
• •Subjective-logic consensus and discount operators fuse independent witness evidence correctly

Generated with AI, which can make mistakes.