Terms
| Term | Meaning |
|---|---|
| Passage | One audited sign-in through the Gate. It is the unit Obelisk meters and bills — you pay per person actually let in, not per stored account. |
| Receipt | An append-only, hash-chained record of a privileged action. Receipts chain to a fixed genesis constant, so a deleted or reordered row is detectable rather than merely unlikely. |
| Assurance | How a session was proved: verified (a passkey), totp, magic link, break-glass, remembered, or federated. Capability is bound to assurance, so a weaker proof cannot perform a sensitive action even with a valid session. |
| Step-up | Asking for a stronger proof before a specific action, rather than blocking the session outright. A magic-link session can read your account and must pass a passkey check to change your credentials. |
| Grant | A short-lived, scoped authorization issued to an agent or an API key. Grants expire on their own; revoking one does not require finding everywhere it was used. |
| DPoP | Demonstrating Proof-of-Possession (RFC 9449). Binds a token to a key the holder must prove control of, so a stolen bearer token alone is not enough to use it. |
| Macaroon | A capability token that can be attenuated — a holder can hand on a strictly weaker version without asking the issuer. Obelisk sessions are macaroons, which is why a session can be narrowed but never widened. |
| Relying party | A project that sends people to the Gate to sign in and receives them back. Registering one is what makes its return URL legitimate. |
| CIMD | Client ID Metadata Documents. Lets an MCP client identify itself by a URL it controls instead of pre-registering, so an agent can connect without a human provisioning step. |
| CAEP | Continuous Access Evaluation Protocol, part of Shared Signals (RFC 8417). When a session is revoked or a credential changes, Obelisk emits a security event token so relying parties learn immediately instead of at the next token refresh. |
| Proof Link | A revocable public page describing an agent's independently-checkable claims — key control, an accountable owner, runtime evidence. Deliberately a vector of bounded claims, never a single verdict about whether something is human. |
| Obelisk Rating | A 0-100 score of an organization's security posture, computed from evidence rather than self-report, and always reported alongside its coverage — how much was actually measured. |
| Coverage | The fraction of a Rating's dimensions that were actually measured. A high score over thin coverage is a weaker claim than a middling score over full coverage. |
| Account Shield | The same idea as the Rating, aimed at one person's own account: passkeys, a backup passkey, recovery codes, an authenticator app, and alerts that can really be delivered. |
| Transparency head | A published commitment to the current state of the receipt ledger. Pin it today, compare it tomorrow: a rewritten history cannot produce the same root. |
| Break-glass | The founder's last-resort access path, deliberately narrow and heavily receipted. It exists so that losing every device is survivable, and it can never perform a step-up-bound action. |
Where these appear
Most of these are load-bearing on the integration docs, the trust model and the security posture. If a word on this site is doing work and is not defined here, that is a documentation bug — tell us.
Machines: the same vocabulary drives /llms.txt and /agents.json.