Acceptance criteria
Observable conditions that a result must satisfy before it is accepted or used.
Use the same words for the same boundaries
Stable definitions make recommendations, controls, and verification expectations easier to inspect.
Agent
A system that selects and performs multiple steps toward an objective, often using tools and intermediate state.
Approved environment
A model and processing environment authorized for the applicable data, retention, access, and tool policy.
Audit trail
A record of inputs, evidence, decisions, approvals, actions, and outcomes sufficient for later review.
Autonomy cap
The highest action or decision level a workflow may use after all rules are applied.
Bounded action
An external action constrained by explicit target, permission, approval, monitoring, rollback, and stop limits.
Citation checking
Verifying that a cited source is authoritative, current enough, and actually supports the attributed claim.
Context
Task-specific information supplied to a model for the current interaction or workflow.
Control
A required workflow condition intended to prevent, detect, limit, or recover from a defined failure.
Data minimization
Using only the data necessary for the task and avoiding unrelated or excessive detail.
Disposition
The methodology outcome: proceed, proceed with controls, redesign, or do not use AI.
Escalation trigger
A defined condition that transfers the task to a person with greater authority or expertise.
Evaluation set
Representative normal, edge, adversarial, and failure cases used to test a workflow repeatedly.
External action
A change to a system, record, message, transaction, or other state outside the model response.
Fit band
A qualitative statement of AI suitability for the assessed task: strong, assisted, limited, or avoid.
Grounding
Connecting generated output to supplied or retrieved evidence that can be inspected.
Hallucination
Generated content that is unsupported, fabricated, or inconsistent with the relevant evidence.
Human check
A review step assigned to a person with the evidence, time, competence, and authority needed for that check.
Inherent risk
The consequence risk of a task before recommended controls are applied.
Least privilege
Granting only the tools, records, fields, actions, and duration necessary for the task.
Model
A learned statistical system that maps input to output; it is one component of a complete workflow.
Pilot
A limited trial with representative cases, explicit acceptance thresholds, monitoring, and stop conditions.
Prompt injection
Instructions in user or retrieved content that attempt to redirect a model or tool-enabled workflow.
Qualified reviewer
A person with task competence, evidence access, error-detection ability, time, and authority to reject or stop.
Residual risk
Risk remaining after controls; the v1 methodology does not estimate or certify it.
Retrieval
Selecting relevant material from an identified corpus before generation or decision support.
Rollback
A tested way to restore prior state after an external action.
Source quality
The authority, relevance, currency, independence, and directness of evidence for a claim.
Source traceability
The ability to connect a material claim to its exact source and observation date.
Stop condition
A deterministic condition that halts a workflow before it continues into unsafe or unreliable behavior.
Verifiability
The degree to which important errors can be detected before the output causes harm.