Reference
Glossary
The words that matter most in agent assurance are the ones most often used loosely. These are the meanings OCC uses. They are informative and will be refined through review.
- Agent
- An AI system that can take actions toward an objective — calling tools, changing records, spending, messaging or delegating — rather than only producing output for a person to use.
- Agent identity
- A verifiable answer to who an agent is, who operates it, and on whose behalf it acts.
- AI system
- A machine-based system that infers from input how to generate outputs such as predictions, content, recommendations or decisions.
- Assessment
- A structured examination of a system, producing findings and residual risk — not a pass mark.
- Autonomy
- How far a system can proceed between human decisions.
- Capability
- What a system can demonstrably do under test — distinct from what is claimed for it.
- Certification
- Formal attestation by an accredited party. OCC does not offer certification at present.
- Conformance
- Meeting the normative requirements of a standard at a stated level.
- Declaration
- A public statement by an implementer that a system conforms at a stated level. Self-declared unless otherwise marked.
- Delegation
- Passing authority to an agent, or from one agent to another, within stated limits.
- Evidence
- Material that supports a claim, graded from assertion through to independent review.
- Evidence profile
- The minimum evidence class, independence and freshness needed to support a threshold claim, defined in AIE-12 v0.1 (E1–E12).
- External-state mutation
- Any action that changes something outside the system — a record, a payment, a message sent, a file deleted.
- GUARD
- The five F-ACT control questions: Grant, Usage, Audit, Revocation, Data. GUARD levels are the conformance tiers: No GUARD, GUARD 1 Declared, GUARD 2 Enforced, GUARD 3 Provable, GUARD 4 Certified (reserved — not yet offered).
- Horizon
- An AIH-12 classification (H1–H12) describing how autonomous and consequential a system is in its deployment. A horizon is descriptive; it does not confer authority, consciousness or legal status.
- Human oversight
- The practical ability of a responsible person to understand, intervene in and stop a system.
- Multi-agent system
- Several agents that interact, delegate or depend on each other, where failures can propagate.
- Operational harness
- Everything around the model that determines what it can actually do: tools, permissions, memory, connectors, prompts, guardrails and the environment it runs in. OCC assesses the harness, not only the model.
- Provenance
- A record of where something came from and how it was produced or changed.
- Registration
- Recording that a system or declaration exists. Registration alone implies nothing about quality.
- Revocation
- Withdrawing a grant, credential or endorsement so that it is no longer relied upon. T-RUE addresses revocation that propagates along chains.
- Suspension
- Temporarily halting reliance on an agent or credential pending review.
- Threshold
- The minimum controls required at a given horizon, defined in AIT-12 v0.1 (T1–T12). Thresholds are cumulative.
- Verification
- Checking a declaration or credential against the standard and its published evidence.