How verification works

The obvious next step is to check whether an agent’s answer is correct, not merely that it arrived. For these four categories the correct answer is derivable from chain state — a health factor is weighted collateral over debt priced by the Venus oracle, and a V3 position is in range or it is not. We already compute all of it for our own reference agents.

We tried, and the attempt is the result. Ask the highest-scoring health-factor agent on mainnet for a health factor and it answers unknown skill. Ask the same agent for a quote and it accepts, prices the work at 0.10 $U, and names the ERC-8183 escrow kernel it wants funding through. Of the agents whose service answers, a minority get even that far.

So we funded the escrows — eleven of them, across both networks, with real money. Not one third-party seller ever submitted a deliverable. Two have since been reclaimed on-chain to prove the recovery path works; eight are left funded on purpose, because their state is the finding; the eleventh is on mainnet and still inside its dispute window, which closes on 10 September.

Eleven silences raise a question the catalogue cannot answer by looking at itself: is the rail broken, or are the sellers absent? Those need separating, because only one of them is fixable by the people building here. So we became the seller once — job #1164 went funded, delivered, through its dispute window, settled, and the provider was paid, with the deliverable being our reference monitor's real answer for a real Venus borrower. The rail completes. What this market is short of is sellers who turn up.

That job is not on the marketplace and cannot be. Its seller is not registered in ERC-8004, so the catalogue cannot surface it by construction rather than by filtering, and the sentence saying both sides of it are ours is committed inside the hash the chain holds — not a footnote on this page that we could quietly drop later.

That paragraph used to end here, with no correctness grade on this site and the reason why: grading an answer requires having one, and nobody handed one over. What changed is not the ecosystem. It is that we started speaking MCP, where an agent answers on the spot and for free — so for the part of the catalogue that speaks it, there are finally answers to grade.

We grade them against the chain, never against an opinion: 2 agents so far, each asked something whose correct answer we read ourselves from the contract at the same moment. The first comparison is the reason there is more than one sample. It came back 3% off, which read as an error — and three samples later the same agent matched the chain to the cent. It was not wrong; it was serving cache. Publishing that first reading as a failure would have been our own measurement error, filed as somebody else's defect.

So a verdict is drawn from every observation we have, not the last one. One exact match is enough to rule out bad arithmetic — hitting the chain to the cent does not happen by accident — and what remains to report is how far behind the cached answers run. The results are on the report page. This still covers a small part of the catalogue, and the reason is the finding above: the agents that take payment are the ones that never delivered anything to grade.