How we verify

A guide that's confidently wrong costs a developer weeks. So every claim carries a mark saying how we know it, and the parts we couldn't confirm are listed, not hidden.

The four evidence levels

What the agent actually tests

Without anyone's credentials, the agent can still check real things:

• That the developer console exists, and that it sits behind sign-in (where you take over).
• The OpenID Connect discovery document: real endpoints, PKCE support and advertised scopes.
• The authorize endpoint, called with a fake client ID. A live endpoint rejects it in a recognisable way.
• That the documentation pages it cites load.

It does not create accounts, sign in, or submit apps for review. Approval timelines therefore come from published provider statements or, where none exist, community reports clearly marked unverified.

Quotes are checked mechanically

When the agent marks a claim as documented, it must include a short verbatim quote from an official page it fetched. After the guide is written, code searches the fetched page text for that quote. If it isn't there, the claim is downgraded to inferred and the guide's verification log says so. Claims marked tested must cite a probe that actually ran.

Curated guides

Library guides were researched by agents against official docs, then checked: each quote was matched against its source page, URLs were probed, and the agent's output for the same provider was compared with the curated version. Each guide's “What we checked” section lists the checks and anything we corrected.