The census is only worth anything if you can check it. This page is the
contract: where values come from, what "verified" means, what a null means, and where
the current limitations are.
Sources
▸Vendor pages only. Every value is extracted from
the API vendor's own documentation, pricing or terms pages — never from third-party
directories, blogs or forums.
▸Per-field evidence. Each non-null field carries the
URL of the exact page that states it. A record whose extraction cites a page we did
not crawl is rejected automatically — fabricated sources cannot ship.
▸Polite crawling. Honest user-agent, low request
rates, and robots.txt respected — pages a site disallows are treated as unavailable
and marked, never fetched. Public docs and pricing pages that plain HTTP can't read (JavaScript-rendered
or bot-manager defaults) are fetched through a standard rendering service
(Firecrawl) — public pages only, never logins or paywalls, capped per domain.
Values are still never guessed: no readable page, no record.
What a null means
▸"not documented" on a record: we crawled the
vendor's pages and found no statement for that field. That is a finding, not a
gap — an agent should know a vendor documents no rate limit.
▸Tracked, no record: -1777 of
25 tracked domains currently block crawling or serve JS-only docs.
They are retried every cycle and never published as guesses.
Verification & QA
▸Re-verification runs the snapshot clock: source
pages are re-fetched and diffed against the stored snapshot, and any changed source
triggers re-extraction — the diffs feed the change feed. Each
record shows its own last_verified date.
▸52 golden assertions guard
hand-audited records (Stripe, GitHub, OpenAI, Slack…): any build in which one of those
verified fields regresses to null fails and cannot deploy.
▸Confidence (high · medium · low) is currently
record-level, assigned at extraction; field-level confidence is on the roadmap.
▸Corrections: every page links a
correction form; reports are reviewed by a human and re-verified
against the vendor source before any change.
Independence
▸Placement, never data. Sponsorship and featured
listings buy disclosed visibility. No sponsor can change a record, a field value, a
confidence score, or who gets listed. See sponsors.
Known limitations
▸Multi-product vendors (AWS, Google, Twilio…) are
currently one record per domain; values are scoped to the primary public API and the
provider→product hierarchy is in progress.
▸Category assignment is automated and being
human-reviewed; expect occasional misfiles until that pass completes.
▸JS-only documentation sites limit coverage for some
vendors (marked in the funnel above).