AI Governance Glossary
The vocabulary of responsible AI agent deployment. From execution warrants to risk tiering.
Execution Warrant
A cryptographic authorization token (HMAC-SHA256 signed) that grants an AI agent permission to perform a specific action within defined scope, time, and resource constraints. Without a valid warrant, the agent cannot execute. Warrants are time-limited, scope-constrained, and tamper-evident.
Risk Tiering
A classification system that assigns every agent action a risk level from T0 (minimal risk, auto-approved) to T3 (critical risk, requires multi-party human approval with justification and rollback plan). Risk tiers determine the approval workflow required before execution.
Intent
A structured declaration of what an AI agent wants to do, submitted before execution. An intent includes the action type, target resource, parameters, and the requesting agent's identity. Intents are evaluated by the policy engine before any action is taken.
Policy Engine
The rule evaluation system that determines whether an agent's intent should be approved, denied, or escalated. Policies are defined as code (JSON/YAML) and can include conditions, scopes, time windows, and escalation rules.
Audit Trail
An immutable, chronological record of every agent action, including the intent, policy evaluation, risk assessment, approval decision, warrant issuance, execution result, and verification. Audit trails in Vienna OS are cryptographically linked and compliance-ready for SOC 2, HIPAA, and SOX.
Agent Fleet Management
Centralized monitoring and governance of multiple AI agents across an organization. Fleet management includes agent registration, health monitoring, policy assignment, and activity dashboards.
Human-in-the-Loop (HITL)
A governance pattern where certain agent actions require explicit human approval before execution. In Vienna OS, T2 and T3 risk-tiered actions require one or more human approvers in the approval chain.
T0 (Tier Zero)
The lowest risk tier. Actions classified as T0 are auto-approved without human intervention. Examples: read-only queries, analytics lookups, status checks. T0 actions still generate audit trail entries.
T1 (Tier One)
Low-risk actions that require policy evaluation but not human approval. If the action matches an approved policy, it proceeds automatically. Examples: deploying to staging, sending notifications to internal teams.
T2 (Tier Two)
Medium-risk actions requiring at least one human approver. Examples: deploying to production, processing payments, modifying customer records. T2 warrants include approval timestamps and approver identity.
T3 (Tier Three)
Critical-risk actions requiring multi-party human approval, a written justification, and a rollback plan. Examples: wire transfers over $50K, bulk data deletion, infrastructure changes affecting multiple services. T3 is the highest governance tier.
Rollback Plan
A required component of T3 warrant requests that specifies how to reverse the action if something goes wrong. The rollback plan is cryptographically linked to the warrant and can be triggered automatically or manually.
Policy-as-Code
The practice of defining governance rules in machine-readable formats (JSON, YAML, or code) rather than human-written documents. Policy-as-code enables automated evaluation, version control, and audit trails for governance decisions.
Warrant Scope
The boundaries defining what a warrant authorizes. Scope includes the action type, target resource, allowed parameters, maximum values, and geographic constraints. Any action outside the warrant scope is denied.
Warrant TTL (Time to Live)
The maximum duration a warrant remains valid after issuance. After the TTL expires, the warrant is automatically invalidated and the agent must request a new one. Default TTLs range from 60 seconds (T3) to 3600 seconds (T0).
Truth Snapshot
A cryptographic hash of the system state at the moment a warrant is issued. During post-execution verification, the truth snapshot is compared to the current state to detect unauthorized modifications.
BSL 1.1 (Business Source License)
A source-available license that allows free use for evaluation, development, and non-production purposes. Production use requires a commercial license. After a specified date (2030 for Vienna OS), the license automatically converts to Apache 2.0.
EU AI Act
European Union regulation establishing a legal framework for artificial intelligence. The EU AI Act classifies AI systems by risk level and imposes requirements for transparency, human oversight, and accountability. Vienna OS helps organizations comply with high-risk AI system requirements.