Security & trust
Sessions, evaluation gates, and the difference between measured, reported, and target.
Remote providers are optional and credential-gated. This website does not see your documents, weights, or generated output from your own ABI runtime. Unknown MCP tools fail closed in that runtime.
A green web check is not mobile evidence. A reported test suite is not an independently reproduced benchmark.
Sign-in is first-party Better Auth: email and password, passkeys, and Google, Apple or X when their credentials are configured. It opens the console, which holds your field notes, a consent-gated model chat, and the audits of those chats. It is not an Abbey session and not a hosted WDBX store.
Chat audits and workspace refresh tokens are sealed with AES-256-GCM under a server key. Without that key those features refuse rather than store plaintext. You can read and delete your audits, and deleting your account removes your per-user rows.
Administrators are an email allowlist plus a linked Google or Apple account. An allowlisted email-and-password account is refused, because sign-up is open and unverified.
Figures on this site are tagged measured, target, or reported. A new number ships only with the source that produced it. Planned is never shipping. Research is not a Quesar product claim.
Further reading linked by the ABI README; not separately audited here.
This site's own auth, admin, encryption and configuration decisions.
Checked against these sources on .