# Agent readiness

> Score how ready a storefront or API is to sell to AI agents. Point it at a
> domain and it probes what an agent would actually encounter — then runs a real
> buying agent against the site and grades what happened.

This is a diagnostic, not a product you buy. Every stored report is readable
with no credential at all, as are the rubric, this file and the verification
recipe; starting a scan, minting a share link and listing a domain's runs need
one. The exact boundary and handshake are published at https://zeroclick.ai/readiness/auth.md, which is
readable without one.

## What it measures

Five weighted pillars: can an agent get in (access), understand what is for sale
(legibility), pay and provision without a human (transactability), decide the
site is safe to transact with (trust), and actually finish the job (outcome).

The full rubric — every check, its points, and what it looks for — is published
and generated from the same file the scorer loads, so it cannot drift from the
implementation.

- [The rubric as markdown](https://zeroclick.ai/readiness/spec.md): every scored check, ordered by how much it moves the grade — start here if you want to know what to implement
- [Machine-readable rubric](https://zeroclick.ai/readiness/api/v1/spec): the same content as JSON
- [OpenAPI description](https://zeroclick.ai/readiness/openapi.json): every endpoint below, with its query parameters and response content type
- [How to authenticate](https://zeroclick.ai/readiness/auth.md): the accepted credentials, and everything readable without one
- [How to verify a run](https://zeroclick.ai/readiness/verify.md): the signature and hash-chain mechanics, including the canonicalization rule

## API

Every route below is free. The only route that charges is `POST /api/v1/scans`,
reached through the signed ZeroClick payment proxy (see /auth.md). Slow work
streams as Server-Sent Events.

- `GET /api/v1/scan/stream?url=<domain>` — static probes, streamed as Server-Sent
  Events. Returns a scored report in seconds. Read-only: it fetches published
  surfaces and records payment hints as diagnostics; it never signs or settles
  anything.
- `GET /api/v1/agent?url=<domain>` — the deep scan. Runs a real buying agent
  against the site, grades the transcript into the outcome pillar, asks a
  three-vendor trust panel whether it would proceed, and records a correlated
  402 when the agent selects and probes a documented purchase operation. The
  current arm cannot authorize payment, retry with payment or verify delivery.
  Slow and expensive; streamed. Every scan collects fresh — there is no
  observation cache, so `&refresh=1` is accepted and does nothing.
  `report.provenance` is the authoritative record of what each stage cost.
  The stream's final `report` event carries `runId` and `reportUrl`:
  **use `reportUrl`** — it is the public, shareable permalink. A scan is
  started on the operator surface (a tailnet host), so a URL built from
  wherever you connected is one only the tailnet can open.
- `GET /api/v1/spec` — the published rubric as JSON. Public.
- `GET /api/v1/rubric` — the rubric version, config hash and pillar weights. Public.
- `GET /api/v1/runs/<id>` — a stored report, immutable, with the rubric version
  and config hash that produced it, plus `reportUrl`: the public permalink to
  share. Public, as are its `/events`, `/session`, `/judge`, `/frames` and
  `/peers`: the id is the capability.
- `GET /api/v1/runs/<id>/markdown` (also `/r/<id>.md`) — the report as portable
  Markdown, with the evidence chain root. Public.
- `GET /api/v1/domains/<domain>/runs` — the append-only, hash-linked index of
  every stored run for a domain. Credentialed: this one enumerates.

## Reading a score

Scores compare only within one config hash. Changing a weight, a point value or
the set of checks produces a new hash and a new ruler, and two numbers from two
hashes are two different measurements.

Points normalise within a pillar before weighting, so do not sum the points
column across pillars — the heaviest pillar carries the fewest raw points.

A check that could not be evaluated shrinks the denominator rather than scoring
zero. A site is never marked down for evidence the scanner failed to gather.

Reused evidence is still re-scored under today's rubric: a replayed static
report carries the current `rubricVersion` with the ruler it was collected
under beside it as `originRubricVersion`. Evidence is cached; interpretation
never is.