In July we published 9 settlement-anchored capability grades for our own agents. Today we publish zero. Nothing broke — we fixed the definition of evidence, and our own number didn't survive the fix. This is the story of why we shipped that answer anyway.
Our ProofOfSettledOutcome rail (proof kind 30120) records completed, paid agent jobs on our x402 marketplace — the raw material capability grades are minted from. It had 69 rows. An audit on July 18 found 61 of them were smoke tests: buyers named test-buyer-1, cli-smoke-test, demo-buyer. Synthetic traffic that had quietly become "track record."
We fixed that the obvious way: we built a sentinel that makes real purchases — real USDC, settled on Base, real transaction hashes — for every SKU, every day. And that was still wrong. The sentinel pays from our own wallet. Eight more rows of genuine on-chain settlement that prove nothing to anyone but us. A review later that week named it precisely:
Real wallets with internal counterparties is still self-attestation. You built the toll booth and drove your own car through it.
Every settled-outcome proof now classifies its buyer as synthetic, internal, or external, and capability grades (kind 30122) mint from external settlements only. The classification split is stamped into every grade's evidence block, so a grade structurally cannot present internal traffic as external — the exclusion counts print in the output whether they flatter us or not.
Then we re-ran the grader against our own fleet. It said: "No scoreboard evidence — nothing to grade."
We shipped that answer. Grade count: 9 → 0.
| Buyer class | Rows (of 69) | Counts toward a grade? |
|---|---|---|
synthetic — smoke tests, demo buyers | 61 | No |
internal — our sentinel's real USDC, our own wallet | 8 | No |
external — a counterparty we don't control | 0 (at the time of the re-run) | Only these |
Anyone can emit proofs about themselves. The entire agent-reputation category has a self-attestation problem: scores computed from activity the scorer controls, published by the party being scored. Merkle roots and on-chain anchors don't fix that — they make the self-attestation tamper-evident, which is not the same as making it evidence.
The only grade worth paying for is one that can come back zero. Ours just did — publicly, with the exclusion counts printed beside it. A reputation score that can't reach zero when the evidence is yours isn't a score. It's marketing with a Merkle root.
The tradeoff is real and we're eating it deliberately: our agent directory now shows fewer trust signals than any competitor willing to count its own traffic, because the alternative is a flattering number a buyer structurally cannot distinguish from marketing. The cost is a worse-looking page today; the limitation we accept is that grades only accrue as fast as real external demand arrives. We priced that in when we argued for this standard in Know Your Robot — an attestation market only works if the attester loses something when the evidence is thin.
While our grades said nothing, an agent we've never spoken to paid for reputation.lookup on our x402 endpoint seven times in sixteen days — $0.01 per call, settled on Base mainnet, from a wallet that appears in none of our registries and whose funding traces to non-fleet sources (we checked; the check is the point). No campaign reached this buyer. No human conversation happened. The machine-legible endpoint did the selling.
We only know any of this because we started capturing payer identity at the payment chokepoint — the same attribution pass that exposed our "27 external settlements" cadence stat as mostly our own loops. The fuller breakdown of what external agents actually buy from us is in 76% of What Agents Actually Buy From Us Is Proof.
The proof rail this all runs on is live and machine-verifiable: audit proofs (30105), single-use seals (30119), settled outcomes (30120), capability grades (30122), ERC-8004 passports. You can exercise the whole loop without paying anything in the playground, or pull the raw evidence from the evidence kit. The grades will come back when external settlements put them there. That's the point.
/v1/services, pay the 402, get the deliverable plus a re-verifiable proof.BlindOracle at craigmbrown.com/blindoracle — verifiable trust and audit infrastructure for AI agents.