Check an x402 endpoint before you pay it.

A: we paid and it delivered, cleanly and reliably. B: we paid and it delivered, with a caveat (the response differs from its listing, a value warning, or 80–95% uptime). C: probably works, but we couldn't confirm it (not paid-tested yet, needs its own sign-in, refused our test wallet, customers only, 50–80% uptime, or an unconfirmed wallet change). D: doesn't work as listed, but you won't lose money (an error after payment with no charge, the listing's own example input gets "not found", or it wants a header or input the listing doesn't document). F: costs you money or is unsafe (took the money and failed, a payout-wallet incident, or under 50% uptime). A and B are proceed; C and D are caution; F is avoid.

Uptime counts how often the endpoint returns a valid payment quote over 7 days. Uptime sets the ceiling: an A needs at least 95% uptime and a passed paid test; under 95% the best grade is B, under 80% it's C, and under 50% it's F and avoid, since it fails most calls. These limits apply once there are at least 6 hours of checks. Latency earns full points up to 1 second and none at 5 seconds; we check from one region, so the curve doesn't mark sellers down for distance. Checks that overlap a stall on our side don't count.

Stability drops when the payout wallet or price changes. Each distinct price increase costs 4 points (8 at most), so prices that move back and forth aren't penalized again and again. An unexplained wallet change is F (avoid), since it can mean the endpoint was hijacked.

Delivery comes from real test payments. If we pay, the money is taken, and we get junk back twice in a row, it's F (avoid) until a paid re-test passes. A gateway error when we pay (HTTP 502, 503, or 504) counts as downtime, not junk. Endpoints not yet paid-tested are scored on the other 80 points, marked "not yet paid-tested", and are C at best: A and B mean a real paid test delivered.

What counts when we pay: any 2xx status is a success. A rate limit or firewall block right after payment, and an async "accepted" (HTTP 202) without the result, are neutral, never failures (if the 202 links to the job, we check the link and grade the finished result). A failure the seller refunds on-chain counts as no money lost (D). An empty result list is an honest "no results". A response fails only when it's an error status, empty when the listing promises output, not JSON when JSON is promised, or has none of the fields the seller's own listing describes. A "sign in first" answer (HTTP 401/403) after payment with no money taken means it needs its own sign-in: C, delivery unconfirmed, with a note (the same for "your wallet isn't allowed" or "customers only"). If the refusal asks for a header or input the listing doesn't mention, that's D: it can't be called as listed until the seller documents it or declares a test request. "Not found" (404) for the listing's own example input, or any other error after payment that doesn't collect the money, is D: it doesn't work as listed, but buyers keep their funds. Taking the money and failing is F (avoid). Each paid test sends a fresh request key, so a seller that de-duplicates requests never sees us as a repeat. An endpoint that failed a paid check is re-tested daily, so a fix shows up quickly.

Listing accuracy: a paid response is compared with the schema or example in the seller's own listing. Data wrapped in a data or result field, or a list of the items the listing describes, still counts as delivered. When the data is reshaped compared with the listing (fields nested elsewhere, a list where the listing shows one item, most example fields missing), the grade is capped at B (85) and the verdict stays proceed, with a note telling agents where to read the fields; matching the listing lifts the cap on the next paid check. Minor gaps (an optional field that isn't returned, a content-type label) get a note but don't affect the grade.

Verdicts: proceed (A, B) means a real paid test delivered and the endpoint is reliable. Healthy endpoints that haven't been paid-tested yet are C, caution (delivery not yet confirmed), which our free checks keep current every 30 minutes; agents can still pay them, or require confirmed delivery.

Value checks: on the paid test we already make, we also check that the answer is plausible: it's about what we asked (the symbol, address or URL we sent), its "generated at" time is recent, and a token's USD price is within 10% of the public spot price. A failed value check is a warning with a B cap, never avoid, since automatic checks can misjudge; the next passing paid test lifts it.

Who gets paid-tested: endpoints with the most real payments first, then those too new to grade, then healthy ones not yet tested. A passed paid test stays valid until something material changes (a price increase, or a new payout wallet or network), so we don't re-pay endpoints just because time passed. Endpoints that failed a paid test are re-tested daily. Our budget covers up to $0.01 per call for any endpoint and up to $0.05 for endpoints agents use or that real payments settle to, with at most 3 endpoints per seller per run.

We never fail an endpoint on one try. A bad response is re-tested about two hours later, and only a second failure counts. If an endpoint rejects our request for missing input, that's on us and never counts.

Buyers' agents can report whether a paid call worked. Reports are tips, not votes: when several independent buyers report a problem, we re-test it ourselves, and the grade only changes if our own paid test confirms it.

A failed payment on our side never counts against a seller.