A principal can’t prove their API works for agents by asserting it. api.qa grades a domain from its own published contracts, deterministically, and signs the verdict — judged by api.qa, never self-graded.
$ curl https://api.qa/auto.dev # api.qa report — auto.dev > Grade A+ · AX 10/10 · remote · attested PASS llms.txt — served and agent-actionable PASS agents.json — capability card parses PASS icp.json — self-classification surface PASS Content negotiation — curl gets markdown, browser gets HTML PASS OpenAPI — machine-readable contract published PASS MCP — interface declared with transport and tools PASS Keyless flow — a No-ask Zone endpoint answers 2xx with no key PASS 402 offers — payment boundaries are structured, hard-ceiling offers PASS Linkset — surfaces cross-reference each other PASS Attestation — identity / attestation ladder declared honesty caps ......... none triggered evidence ............. 11 exchanges re-judge ............. exit 0
How it works
Six mechanisms. Each is labelled with the attack from the threat model it exists to defeat — the number is the attack, not a sequence.
Checks are computed from the target’s own published surfaces. There are zero repo-local test files to rewrite until green.
Acceptance names a sha256 digest, not a file path. If the spec text doesn’t hash to the pin, nothing runs — the verifier refuses before a single probe fires.
Attested verdicts are Ed25519-signed by a key that lives only as a deploy secret. A fleet that owns the code still cannot mint attested history.
The evidence bundle is embedded, so anyone can re-judge it offline and confirm the grade reproduces. The same evidence cannot judge two ways.
Endpoint sampling is seeded fresh per run and recorded in the report. Overfitting to one run’s probe set fails the next.
Two non-scoring checks cap the grade at C when a surface lies. A faked 200 where a typed BLOCKED belongs scores worse than a missing surface.
The AX score
Each item is derived from the target’s own published surfaces and scored 0 or 1, over the machine surfaces made normative by AXP — the Agent eXperience Protocol (https://apis.ax/axp).
A verdict is a pure function of five inputs. None of them is yours to write.
The villain is the human-first web — surfaces built for eyes, gated by signups, that lie to machines (walls of HTML, faked 200s, prices discovered after the bill). A fleet that would rather game the grade than fix the product finds the tests held outside its write access.
Two heroes, one motion
grade and attested, and integrate only on a verdict that clears your bar.passed: true at a ratified digest, or nothing, from a service the fleet cannot write to.Pricing
The public grade runs unauthenticated, because a gate on the free grade would contradict the thesis. Money enters only at the boundaries, as machine-settleable 402 offers.
The stakes
Fail, and the agent-first web routes around you: to agents you are invisible, to wallets you are unbounded risk. Pass, and an agent that has never seen you discovers, understands, and pays you on first contact.