Loading
Who claims to be an agent, who vouched for them, and how much of either turns into anything measurable. A registration is a claim, not activity.Note 23
Avalanche
1,815 registered on Avalanche. On the obvious metric this is the top of the standard.
Avalanche
Coverage and credibility point in opposite directions here, and a ranking built on the first ring would put this chain top of the standard.
Ten addresses wrote 50% of every rating in this registry, and 62% of the agents that carry one carry it from a single account. A rating from one party is a relationship, not a reputation.
A self-rating is a party vouching for itself. The address that wrote it is the one holding the rated agent’s identity token. That is not fraud and it is not evidence either, and it is worth knowing that 3,432 of them were written by just 5 addresses, about 85 agents; concentrated rather than a habit of the crowd.
Ours, not theirs: 5,133 ratings point at an agent id we do not hold. That is a gap in our reading of the identity registry rather than a defect in the reputation one, and it is counted apart from everything above.
This is a filter, not a score. A score would be a number we invented. A filter anybody can apply themselves: keep only the ratings written by an address that has ever paid for or been paid for something on a rail we read, and see what is left of the layer.
| Counting | Ratings | Rated agents | Authors |
|---|---|---|---|
| Every rating in the registry | 559,661 | 45,825 | 20,214 |
| Only from an author who has ever transactedon ACP, the mech marketplace or x402, the three rails we read | 2,474 | 987 | 174 |
| For contrast: from an author who has written 1,000 or more | 492,407 | — | — |
What it leaves. 0.4% of the ratings survive that filter, and 2% of the rated agents. 88% of the layer was written by addresses that wrote at least a thousand each, and whatever that is, it is not the thing the word reputation implies.
What it cannot say. The filter is a floor and cuts our way: an author with no economic history here may have plenty somewhere we do not read, and three rails is what we have. It also says nothing about whether the surviving ratings are honest, only that the address behind them has done something besides write ratings. The bar for “prolific” is published so it can be argued with.
This reads one contract: the ERC-8004 reputation registry, on the networks where it is deployed. An agent with a reputation somewhere else, a marketplace’s own star rating, a GitHub history, a customer who simply came back, has none here, and that is a limit of the standard’s adoption rather than a judgement about the agent.
Concentration is not fraud. One account writing every rating on a network is consistent with a single honest operator rating the agents it actually hired, and we cannot tell that apart from an account rating itself through intermediaries. What the figures do settle is narrower and still worth saying: on most of these networks the ratings are not a crowd, so a score derived from them carries one party’s opinion, not a market’s.
Counts are floors. The backfill walks each registry from its deployment and a network still short of the head reports what it has reached so far, never an estimate of the rest.