Policy & Authority

Technical access is not institutional authority.

The AI Governance and Evidence Infrastructure (AGEI) establishes a rigid boundary between what an entity is technically capable of accessing (APIs, databases, network connections) and the institutional authority it has been granted to perform a specific action.

Machine-Evaluable Policy Rules

Policies are not documents sitting on a SharePoint drive. They are versioned, machine-readable artifacts. The schema tracks policy_sets, policy_versions, and policy_rules to convert governance narratives into exact, evaluable logic constraints.

Delegated Authority and Attenuation

Under the Proof-Carrying Dynamic Authority Model, an agent never operates with ambient, root-level system power. It must borrow bounded authority from a delegating principal (human or parent agent). Furthermore, delegation cannot create or amplify privileges; it strictly attenuates them.

Just-In-Time (JIT) Elevation

By default, high-risk tools and actions are denied. An agent must request Just-In-Time privilege elevation, which often involves a gate evaluation requiring a secondary human approval signal. This produces a cryptographically signed approval ticket stored within the agent's proof bundle.

Key Authority Concepts

  • Purpose Limitation: Authority is granted for a specific contextual purpose, enforcing strict boundaries on how data and tools can be combined.
  • ABAC/RBAC and Resource Scope: Traditional Role-Based Access Control and Attribute-Based Access Control restrict which resources an agent's delegated token can even theoretically touch.
  • Approval Workflows and Separation of Duties (SoD): "Escalating" an action generates a durable, attributable decision trail (e.g., approval_requests). The person who initiates a task cannot be the same person who approves the JIT exception.
  • Time, Spend, and Action Limits: Delegated authority tokens specify not_before, expires_at, and single_use_nonce fields to ensure temporary grants cannot be reused or hoarded.
  • Policy Versioning and Exception Governance: Overriding a policy requires an explicit, audited exception request, binding the exact rule bypassed to the human principal who authorized it.