AxonBrief Academy

Acceptance criteria

Conditions, measures, and thresholds specified before evaluation that determine whether a workflow is fit for a defined use and release scope.

Accountability

The obligation of an actor to answer for decisions, actions, or outcomes within its role and authority.

Agent control plane

An independent runtime layer that converts organizational identity, risk, policy, approvals, and delegation into bounded machine authority and produces evidence that those boundaries were enforced.

AI actor / stakeholder

A person or organization participating in, responsible for, affected by, or connected to AI-system activities. Consult ISO/IEC 22989 and OECD sources for formal wording.

AI lifecycle

The stages through which an AI system progresses, including conception, design, development, evaluation, deployment, operation, change, and retirement.

AI system

An engineered system that generates outputs such as content, forecasts, recommendations, or decisions for given objectives. Consult ISO/IEC 22989 for standardized terminology.

Authoritative source

A source designated by organizational policy as the system of record or source of truth for a particular decision domain.

Baseline

Measured performance of the current workflow on comparable cases, used to assess whether the proposed AI system improves the complete job.

Bounded autonomy

Autonomous system behavior operating within explicit, enforceable, observable, and revocable authority boundaries.

Capability omission

Removing or withholding a capability that should never be used rather than relying only on instructions telling the system not to use it.

Compensating control

An alternative control used to reduce risk when a preferred control cannot be met temporarily or in a particular context.

Context pipeline

The sequence that selects, checks, assembles, and traces task-relevant information from source systems into the context used by an AI workflow.

Continuous assurance

Ongoing collection and evaluation of evidence that governance conditions and control effectiveness remain valid as the system and environment change.

Control objective

A testable condition derived from governance intent and policy that must remain true for a system to be appropriately governed.

Cost per trusted outcome

Total cost of producing a completed result that meets quality and boundary requirements, including model calls, infrastructure, review, and rework.

Decision right

Explicitly allocated authority to make a defined class of decisions or authorize actions under stated conditions and limits.

Delegated authority

Authority transferred by a legitimate actor or system to another actor or system for a defined task and bounded scope.

Escalation

A governance outcome in which a decision cannot be automatically permitted or denied and is transferred to an actor with higher or different authority.

Evaluation gate

A lifecycle point at which defined evaluation evidence is checked before a system may proceed to production or expanded authority.

Evaluation governance

Governance of what is evaluated, admissible metrics and test sets, threshold authority, interpretation, and sufficient evidence.

Exception

A temporary, bounded, formally authorized deviation from normal policy, with scope, duration, compensating controls, evidence, and expiry.

Execution capability

The technical ability of a person, service, agent, credential, or component to perform an action. Capability is not proof of legitimate authority.

Execution envelope

The bounded identities, actions, resources, data, tools, limits, duration, context, delegation, and escalation rules defining legitimate machine authority.

Explainability

The degree to which meaningful information can be provided about why an AI system generated an output or behaved in a particular way.

Fallback

A defined alternate path when a dependency fails, such as manual review or another approved provider, that preserves the workflow's quality and authority constraints.

Freshness

Whether information is sufficiently current for the task, considering source effective dates, indexing delay, expiry, and the time of use.

Governance drift

Gradual divergence between the live system and the authority, controls, evaluations, or assumptions under which it was approved.

Governance evidence

Trustworthy information sufficient to show that a decision, control, or authority boundary was legitimately established and operated as intended.

Governance invariant

A control objective expressed as a condition that remains true despite normal variability in model behavior.

Governed AI registry

An inventory covering use cases, owners, risk, data/context, memory, tools, credentials, providers, actions, evaluations, exceptions, and lifecycle—not only models.

Hard boundary

A condition that must not be violated, such as a forbidden business effect or data disclosure; it is evaluated separately from average answer quality.

Holdout set

Representative evaluation cases kept outside prompt or system tuning to provide a less biased check of changes.

Human oversight

Mechanisms through which humans supervise, approve, intervene in, override, suspend, or otherwise govern AI-system activity.

Idempotency

A property of an operation that lets a retry produce the same intended effect without duplicating a consequential action.

Impact

A consequence affecting people, groups, organizations, rights, processes, assets, society, or the environment. Impact assessment is broader than a numeric risk tier.

Independent challenge

Review by an actor or control function sufficiently independent from the party with incentives to build, ship, or operate the system.

Inference

Use of a trained model to produce outputs from inputs.

Least authority

Granting only the minimum authority required for the current task, context, and duration.

Machine authority

The formally delegated capacity of an AI-enabled system to change organizational or external state within explicit boundaries.

Meaningful human oversight

Oversight in which a competent person has legitimate authority, sufficient information and time, and an effective ability to intervene, reject, escalate, or stop.

Model

A computational representation used to generate predictions, content, classifications, decisions, or other outputs from inputs.

Model capability

What a model can generate, classify, or reason about when given appropriate inputs; this alone does not establish end-to-end system performance or authority.

Model portfolio

A set of eligible models used for different tasks or conditions, selected through evaluation of quality, data constraints, reliability, latency, and total cost.

Model-routing governance

Constraints that determine which models and providers are admissible before optimization for cost, latency, or quality.

Non-amplifying delegation

Delegated authority cannot exceed that of the delegating chain and should normally be attenuated to the minimum required.

Observability

The ability to understand a live workflow's behavior and outcomes through traces, metrics, events, and evidence across its components.

Policy Decision Point

A component or service that evaluates relevant facts against policy and returns an authorization decision.

Policy Enforcement Point

A component or service that mediates access and enforces a policy decision before execution.

Policy-as-code

Compiling selected explicit governance decisions into deterministic, versioned, testable rules evaluated at authoritative control points.

Provenance

Information about the origin and lineage of data, context, models, artifacts, decisions, approvals, or other elements relevant to behavior.

Readiness review

An evidence-based decision about whether a specific system may enter a defined pilot or release scope, including gaps, owners, limits, and stop conditions.

Release envelope

The users, cases, data, actions, models, volume, duration, monitoring, and stop conditions authorized for a particular release.

Residual risk

Risk remaining after controls and treatment have been applied.

Retrieval eligibility

The rule that a source may enter the workflow's context only if it is accessible, applicable, and permitted for the user and task.

Reversibility

The degree to which consequences can be undone or remediated. Lower reversibility generally warrants stronger authority and oversight controls.

Risk

The effect of uncertainty on objectives; AI risk analysis considers likelihood, consequence, affected parties, context, and controls.

Risk-acceptance authority

Legitimate authority to accept a defined level and type of residual risk on behalf of the organization.

Segregation of duties

Separation of proposing, approving, executing, and verifying so one actor does not control the full critical chain without challenge.

Shadow operation

Running an AI workflow on live inputs without letting its outputs affect real decisions, to observe data and integration behavior before a pilot.

System capability

What the assembled workflow reliably accomplishes within its actual data, authority, latency, cost, and operational constraints.

Tool contract

The specification of a tool's inputs, allowed caller and resources, preconditions, effects, outputs, errors, timeouts, and evidence.

Traceability

Ability to link events, artifacts, decisions, components, and lifecycle records so evolution and operation can be reconstructed.

Training

The process through which model parameters are adjusted using data and an optimization procedure.

Transparency

Availability of appropriate information about an AI system, its operation, limitations, or governance to relevant stakeholders.

AI governance and systems engineering vocabulary.