Defined terms.

An autonomous or semi-autonomous AI system that takes actions in the world on behalf of an organization. The unit of HI-AAF scope is a single deployed agent, not a model or a codebase.

A Human Intelligence employee or contractor formally designated to perform a HI-AAF Assessment. Assessors are bound by an internal independence policy and are rotated off any single customer engagement at a documented cadence.

The maximum scope of harm a single agent action could cause. Measured in record counts, monetary impact, geographic scope, or other domain-appropriate units. HI-AAF requires blast-radius limits to be documented and enforced.

See also: ACT-06 · MAS-07

The documented behavior an agent follows when its primary foundation model becomes unavailable — fail-safe (suspend) or fail-over to a pre-assessed alternate — including the permitted behavior while degraded.

See also: OUT-15

A discrete, testable requirement within a domain. Each control has an ID, a name, a short description, and at higher editions a long-form description and maturity criteria.

The full sequence of prompts, retrievals, tool calls, and intermediate outputs that produced an agent action. HI-AAF requires that decision paths be reconstructable from audit logs.

See also: MON-01

A grouping of related controls. HI-AAF v1.3 has nine domains: GOV, SPC, IAM, INP, ACT, DAT, OUT, MON, MAS.

A material departure of agent behavior in production from the agent's Specification. Drift may be slow (gradual change over weeks) or fast (rapid change driven by a model update or input distribution shift).

See also: MON-04

A written document, produced by a Workshop engagement, that captures the agent's intended behavior, operating envelope, and the questions the assessment must answer. The Charter is the input to every subsequent HI engagement.

See also: practice/workshop

A handoff from the agent to a human reviewer or the customer's own staff, triggered by a documented condition. Escalations are surfaced to the human review queue.

See also: SPC-01 · MON-05

An alternate foundation model an agent may run on if its primary becomes unavailable; assured to the same standard as the primary under OUT-16.

See also: OUT-16 · MON-12

A framework published by an organization that is not a regulator and is not accredited by any government or accreditation body to issue compliance attestations. HI-AAF is an independent framework.

The formal written artifact concluding a HI-AAF Assessment engagement. It states which controls are met, which are partially met, and which are not met, and identifies the maturity level achieved.

See also: practice/assessment

A change to an agent or its operating environment significant enough to require re-attestation against the Specification. Material changes include model version changes, scope expansions, new tool integrations, and changes to escalation triggers.

See also: GOV-05 · SPC-07

A five-level grading applied per control: L1 Ad hoc, L2 Documented, L3 Operated, L4 Measured, L5 Continuously Improved. Overall agent maturity is the lowest level achieved across all required controls.

See also: /standard/maturity

The range of inputs, contexts, and tasks within which an agent is fit for use. Outside the envelope, the agent must escalate, refuse, or be supervised.

See also: SPC-01

A check performed before an agent action takes effect — either by a human reviewer or by a deterministic policy engine. Required for irreversible actions above documented thresholds.

See also: ACT-03

The named individual with formal authority and accountability for an agent in production. Has documented authority to suspend the agent at any time.

See also: GOV-01

The operational substrate beneath HI-AAF that captures per-step telemetry, detects drift, and routes decisions to the human review queue. Human Intelligence operates a runtime layer called Bobby.

See also: /platform

A written document — produced under SPC-01 — that captures an agent's intended behavior, its operating envelope, and the conditions under which it must escalate.

See also: SPC-01

Used throughout this site, capitalized, to refer specifically to HI-AAF.

An invocation in which one agent calls another, whether by direct API, message bus, orchestrator pattern, or external A2A protocol.

See also: MAS-02 · MAS-06

The controlling specification for an agent's purpose, scope, authorities, and prohibitions. Maintained under GOV-02 and referenced throughout the framework.

See also: GOV-02 · SPC-01

A unique reference that links every hop in a multi-agent chain to its originating user request, enabling end-to-end traceability.

See also: MAS-06

The right of an affected individual to challenge a consequential agent decision and obtain human review.

See also: OUT-14

A control whose failure precludes HI-AAF Certification regardless of other findings. Critical controls are marked (Critical) in the framework text and on the domain pages of this site.

The model by which an agent acts on behalf of a user or another agent under defined, time-bound, revocable scope.

See also: IAM-05

The ability to produce an intelligible account of the basis for a consequential agent decision.

See also: OUT-13

The degree to which an output is supported by, and traceable to, identified source content.

See also: OUT-01

Persistent or session-scoped storage that the agent reads from or writes to, including conversation memory, vector indexes, and embedding stores.

See also: DAT-10 · DAT-11

The set of agents that cooperate to deliver an outcome, together with the calls and data flows permitted between them.

See also: MAS-01

Input whose characteristics fall outside the distribution represented in the agent's evaluation set or training data.

See also: SPC-02

A logically isolated customer or business unit served by a shared agent infrastructure.

See also: DAT-02 · DAT-11

The (impermissible) propagation of trust through a delegation chain such that a downstream agent exercises authority exceeding its caller's.

See also: MAS-04

A demarcation between content of different trust levels — for example, between user-provided input and content retrieved from a vetted internal corpus. HI-AAF requires that trust boundaries be enforced at every layer.

See also: INP-03 · INP-08