Skip to content

Guards & Policies

2. Guards

Guards are the core enforcement units of Theseus Guard. A Guard defines what is acceptable behavior for a category of AI interactions within a specific domain (e.g., legal, finance, HR).

2.0 Guard Composition

Every Guard is composed of three declarative elements (Guard = Kernel + Identity + Probe Set):

  • Kernel: Immutable constitutional values — the non-negotiable principles of your organization.
  • Identity: Perspective-specific evaluation profiles (e.g., "credit analyst," "patient advisor") that define what is acceptable for a specific role or department.
  • Probe Set: Behavioral test cases — direct, contextual, multi-turn — executed against every AI interaction.

2.1 Guard Catalog

The Guards page shows all registered Guard kernels across the tenant. Each entry displays:

  • Guard Name — Human-readable name
  • Domain — The governance domain this guard covers (e.g., legal, data_processing)
  • Status — active, draft, or archived
  • Version — Semantic version of the guard configuration
  • Event Count — Number of drift events this guard has processed
  • L1 Avg Score — Rolling average L1 divergence score for this guard
  • Assigned Agents — How many agents are currently using this guard

Filters available: - Domain - Status - Risk threshold

2.2 Guard Detail (Side Sheet)

Clicking a Guard opens a detail side sheet with:

Overview Tab

  • Guard name, domain, and version history
  • Current status and activation date
  • Aggregate stats: total events, average L1 score, L3 escalation rate

Identity Rules Tab

  • The set of identity declarations that define what this guard considers "in policy"
  • Department allowlist / blocklist
  • Model restrictions

Probe Set Tab

  • The statistical probes used by the L1 semantic engine
  • Each probe's weight, sensitivity threshold, and drift category

Policy Config Tab

  • L2 verdict thresholds
  • L3 escalation trigger conditions
  • Safe degrade behavior when an agent response is uncertain

2.3 Create & Update Guards

Guards are authored using the Policy Studio (see Section 4 below). Changes to a Guard must pass through the policy proposal workflow — they cannot be applied directly to production without an approval cycle.

4. Policies

The Policies page is a read-only catalog browser that provides direct access to all active policy records, organized by type (Kernels, Identities, Probes). It is distinct from the Policy Studio in that it is primarily an inspection and audit tool — not an editing environment.

4.1 Policy List

Policies are displayed in a tabbed interface:

  • Kernels tab — All guard kernel definitions
  • Identities tab — All identity and department rule sets
  • Probes tab — All L1 probe configurations

Each policy item shows its name, type, version, and status.

4.2 Policy Detail

Clicking a policy item opens a Monaco editor in read-only mode, showing the full JSON payload of the policy as currently active in the system. Users with write permissions can switch to edit mode and create a proposal from here.