Start with the strongest available source.
We prefer official specifications and repositories, developer documentation, product/pricing documents and official announcements. Independent measurements can establish observed behavior. A vendor’s performance or revenue claim remains a vendor claim. Community signals can trigger research, but do not alone justify a hard merchant recommendation.
A claim always has a scope.
Material observations keep a source ID, URL, checked date, capture date, evidence class and region/plan context where known. A released specification is not evidence of a particular merchant deployment. Unknown does not mean no, and an inaccessible source does not prove a product is unavailable.
The L0 scanner reads public evidence.
It reads robots first, then the homepage, a sitemap and at most one child sitemap, up to three product pages, up to three policy pages and a version-pinned UCP discovery signal. It never runs JavaScript, logs in, submits a form, visits cart/checkout action paths or performs a purchase. It does not probe an invented ACP endpoint.
Requests identify AgenticCommerceWikiScanner/1.0. URLs and DNS answers are verified as public; connections use the verified address. Redirects are checked again. Private and metadata destinations, credentials, nonstandard ports, dangerous schemes, oversized responses and unsafe MIME types are rejected.
Limits are part of the result.
The scan permits at most 18 requests including robots and redirects, 512 KB per response, 2 MB total response data, three redirects per fetch, a five-second request timeout and a 40-second scan budget. Requests are spaced by at least 800 ms and honor crawl delays within the total budget. Compressed responses and cross-site redirects are conservatively excluded.
Deployed scans require a shared atomic limiter for requester, burst, merchant cooldown and daily merchant quotas. Backend errors stop automatic scanning safely; local pasted-evidence analysis remains available. A blocked or unreadable page leaves affected checks unverified; it does not create a negative claim about the merchant.
No composite readiness score.
Verified establishes one expected signal. Needs attention identifies a specific observed gap. Cannot verify publicly preserves uncertainty. Not applicable requires explicit context. Blocked records access restrictions. Every finding includes evidence, source, date, why it matters, a suggested fix and a limitation.
History is append-only.
Our first house baseline is September 2026. Earlier effective dates are recorded retrospectively and never backdated as earlier house observations. Market changes, evidence improvements, scope clarifications and conflicts are different event types. Material changes need editorial review before publication.
Research is bounded.
The initial pilot uses 25 public merchants as a convenience sample. It respects robots, uses a low request rate, makes no purchases and separates access failures from observable product gaps. Results cannot represent the whole market. We publish aggregate findings only; private individual scan reports are not benchmark pages.