Learn

The words, without the jargon.

This category is new enough that the same word means three things depending on who is speaking. These are the definitions this product uses, and where an industry term is genuinely contested we say so rather than picking one and hoping.

Authority

Grant: A statement that one named agent may do specific things, up to a value, until a time, under conditions — signed for by an accountable human. Not a role and not an API key: those say who you are, a grant says what you may do.
Attenuation: The rule that a delegated grant cannot exceed its parent on any dimension. Authority subdivides; it never multiplies.
Enforcement mode: Which rung of the ladder a grant is on. observe and shadow refuse nothing; canary and enforce do. A grant in observe is instrumentation, and calling it a control is the most common self-deception in this field.
Capability token: A short-lived credential bound to one resource, action and amount, which the resource verifies offline. It is what makes a refusal independent of the agent's cooperation.

Evidence

Trajectory: One episode of an agent's work, from goal to outcome, with the steps in between. The governable unit — smaller than a session, larger than a call.
Taint: The record that an agent read something untrusted. A grant can require a clean context, so what an agent read changes what it may do. Borrowed from information-flow control, where it is decades old.
Inclusion proof: A short cryptographic path showing one record is in a sealed batch, without disclosing the batch. It is how a regulator checks one trajectory with no access to your system and no sight of anybody else's.
Observed / asserted / not recorded: Three states, never two. observed means the platform watched it happen; asserted means somebody said so; not_recorded means nobody said. A system that promotes the second to the first is manufacturing evidence.

Regulation

Clause map: One regulatory instrument, written as data: each clause, what evidence would satisfy it, and whether that evidence is about the platform's behaviour or the organisation's process. Data rather than code, so the twelfth instrument costs what the first did.
Market profile: Which laws a workload is under, resolved per decision — narrowest wins — and sealed into the record. Not the same as residency, which is where the data sits, and not the same as locale.
Readiness: How much of an instrument a deployment can currently evidence. Explicitly not a compliance score: it counts clauses that describe the platform's behaviour, over obligations that currently apply.

A term we avoid: "AI guardrails". It is used for content filtering, output validation, prompt-injection defence and authorisation interchangeably, which makes it useless in a sentence where precision matters. This site says which one it means.