What Firsthand is

Firsthand is a memory-native trust layer for agent commerce. It watches real interactions on Base, keeps an isolated record per counterparty, and answers one question: has this counterparty done what it said it would before?

Every answer carries the observations it rests on. Nothing is asserted without showing what it stands on.

The problem it answers

On Base, 90.6% of ERC-8004 reviewers show coordinated Sybil behaviour, 86.8% of rated agents have no valid feedback once it is stripped, and a reputation can be moved for a median of $0.0027. The ERC-8004 specification defers the fix to a reviewer reputation layer that, in its own words, does not yet exist.

Source: Xiong et al., arXiv:2606.26028, independently verifiable.

Firsthand is that layer. It scores claimants by whether anyone else independently witnessed the same agent, and grounds every verdict in an event it read from the chain itself.

What it is not

Firsthand does not compute a reputation score, a rating, or a star count. Those are the broken things it replaces, and using their vocabulary concedes the argument. It returns a standing, a confidence and a basis.

It also never asks a language model what it thinks. The verdict is arithmetic over the record. A sentence of English may be rendered from a verdict for a human reader, but the decision is computed, and running it twice over the same record returns the same answer.

Where to go next