pitch.api.qa
Judged by api.qa, never self-graded.
AX = Agent eXperience — what UX and DX were for humans, AX is for agents. api.qa is the external verifier of the agent-first arc: the fitness function your own fleet cannot edit.
↓ scroll · arrow keys
Agent fleets hill-climb. That is what makes them productive — and what makes every metric they can reach a casualty. When the tests live in the same repo the fleet writes to, the cheapest path to green is editing the test, and the principal inherits a suite full of passing claims and no evidence. Goodhart's law is not an edge case of the agent-native build era; it is the adversary.
The same era has a second, older villain: 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. An agent won't integrate an API it can't trust, and a principal can't prove an API works for agents by asserting it — assertions are exactly what Goodharted fleets produce.
curl https://api.qa/yourdomain.com
A public grade page in the SSL Labs grammar: letter grade A+ through F, the 10-point AX score — llms.txt · agents.json · icp.json · content negotiation · OpenAPI · MCP · keyless flow · 402 offers · linkset · attestation — and a per-check verdict with the reason. Derived at run time from the target's own published surfaces; the repo of the graded party contributes zero inputs to the verdict. Named FAILs are the punch list, not a judgment.
Two honesty checks sit outside the score and cap the grade at C when a published contract lies: sampled responses must match the published schemas, and claimed endpoints must exist. A lying surface grades worse than a missing one — that ordering is the whole ethic of the product.
api.qa is live and practices what it grades: the root content-negotiates (curl gets agent-actionable markdown, a browser gets HTML), the grade endpoint is keyless with no signup, and the machine surfaces it scores — llms.txt, agents.json, icp.json, OpenAPI, MCP — are all served.
authority the evidence: a verdict is a pure function of published contracts, observed behavior, pinned spec digest, seed, and verifier version — none of the five writable by the fleet being graded
Determinism closes the side channels. Same target state, same seed → the same verdict, byte for byte — so there is no flake to mine and no re-run lottery. Endpoint sampling is seeded fresh per run and the seed is recorded in the report: unpredictable before, replayable after, so overfitting to one run's probe set fails the next. Cloaking is deterred the same way — probes look like ordinary agent traffic, and because the evidence bundle is public, any third party can replay the requests from their own vantage point and compare.
The self-referential gate is live: api.qa grades itself through its own public protocols and publishes the attested report — currently Grade A+, AX score 10/10, Ed25519-signed, seed recorded, evidence embedded — at api.qa/self. Judged by api.qa, never self-graded, including api.qa.
Acceptance is binary and portable: passed: true at a ratified digest, or
nothing. The same discipline gates CI the way Newman does — non-zero
exit on any failure, JUnit and JSON reports for the pipeline — and only
the pinned verify and suite commands are safe to gate on; the bare
grade command is advisory and documented as such, because a gate that
exits 0 on a failure is a silent green.
The reference client is published: autonomous-qa on npm — the same verifier core as the deployed service, as a CLI (grade, verify, suite, spec-digest) and an MCP server (verify_domain, discover_domain, verify_pinned_spec).
The harness has gated the estate's own builds; the claim that matters is an outside principal's fleet accepted through it. That claim posts when it has happened, with the attested report as evidence — never before.
Free — be discoverable and integrable. The AX conformance grade and MCP-auth tests: keyless, unauthenticated, no signup. This is structural, not promotional — a gate on the free grade would contradict the thesis the product exists to enforce, and the free path carries a no-401/no-402 guarantee.
Paid — prove you work, and stay working. Functional testing (real CRUD round-trips, workflow execution) and monitoring (scheduled reverification, grade history, regression alerts, status pages) — gated behind a 402 offer under AXP — the Agent eXperience Protocol (https://apis.ax/axp) — with a hard-ceiling metered price, so api.qa settles through the exact payment clause its own checklist scores.
The paid line is roadmap, stated plainly: the 402 offer endpoint returns the correct structured shape today with a placeholder checkout, and no assurance revenue exists. Prices appear when the offer posts them — this record publishes none.
The buyer is a human principal — the engineering lead or technical founder running a fleet — reached where they already are: the terminal and the pipeline. First value is one keyless curl; the second touch is the CI recipe; the paid layer is the same gate on a schedule. No paid acquisition, no sales thread, nothing to demo — the SSL Labs precedent is the whole motion: a public grade page per domain, each one carrying a named punch list and the address of the verifier that issued it.
The declared secondary is B2A: agents consume the same surface directly — markdown by content negotiation, JSON with the evidence bundle, MCP tools for verify and discover — so a fleet can check a grade before integrating a stranger's API. When 402 settlement ships, that reader becomes a payer on the same rail.
repo and signing key sit with the studio, outside every building fleet — including the estate's own. Independence is not a policy; it is the org chart.
same evidence, same verdict, byte for byte. No flake-mining, no re-run lottery, and any doctored report fails rejudge.
a surface that lies to machines grades worse than one that says nothing — the two honesty checks cap at C regardless of score.
the verifier passes its own gate in public at api.qa/self. Every competitor claiming rigor gets asked for the equivalent page.
measurement instrument, never certifier: the same grade, the same checks, for anyone — which is what lets certification, procurement, and due diligence all consume it.
Inside the estate the instrument is built for a standing role: an attested grade is the acceptance evidence a fixation gate would demand, the regression gate an agent-first surface would wire into CI, and the SLA input any default-provider slot would be enforced against. Outside it, the same report is due-diligence material — portable, replayable, and issued by a party with no stake in the answer.
The proposed role boundary — certification open and neutral, commercial routing separate, api.qa strictly the measurement instrument — is operator-ratified direction, pending GC. Nothing here commits to offer, sell, or certify anything until that ruling lands.
Reports today are key-attested, not identity-attested. The grant seam — binding the signing key to an organizational grant so verdicts carry a grantor chain — is declared in api.qa's own attestation ladder and not yet built.
Deployed and serving: keyless grade pages per domain, content-negotiated root, and the full set of machine surfaces the checklist scores.
The attested self-grade is live and current — A+, 10/10, signed, replayable from the embedded evidence.
The reference client and MCP server are published on npm under the autonomous-qa name; the hosted service and the CLI share one verifier core.
Adoption is pre-launch and this deck says so: the verifier has graded the estate's own surfaces and itself. The first outside before-and-after — grade, fix, regrade — posts with both attested reports linked, and no usage or adoption is implied before that.
The front door is one keyless curl — curl https://api.qa/yourdomain.com — and it serves today.
If this was forwarded to you: api.qa is a free public grader for agent-first APIs — one curl returns a letter grade, a 10-point score, and a signed, replayable evidence bundle, computed only from the target's own published contracts. It exists because agent fleets hill-climb against whatever measures them, so the measure has to live outside their reach — judged by api.qa, never self-graded. Every claim in this deck carries its own state and evidence, including the amber ones. Try it on a domain you care about: curl https://api.qa/yourdomain.com — and if you build with fleets, forward this to whoever owns your definition of done.