AI that recommends inside the rules.

• Evidence-linked analysisFindings are built from Reality Graph records and link back to each of them.
• Governed AI assistanceModels and agents act as principals with identity, delegation and limits.
• Classification-aware model routingData reaches a model only through a route cleared for its labels.
• Analyst workflowsCollection, triage, correlation, assessment and review as recorded steps.
• Source citationEvery generated statement cites the records it depends on.
• Uncertainty disclosureConfidence is shown with its method, and alternatives are stated.
• Contradiction detectionConflicts between sources are surfaced, not averaged away.
• Recommendation generationProposed courses of action, marked as proposals with no authority.
• Human reviewAssessments are made and signed by accountable people.

AI in Verity Command is a governed principal, never an autonomous sovereign authority. An agent has an identity, a registered definition, a delegating person and a mission scope. Its effective permissions are the intersection of its own grant, its delegator's current authority and the mission — never more than any of them.

• An AI agent cannot approve, authorize or trigger a high-consequence action.
• An AI agent cannot present an inference as an observation.
• An AI agent loses its permissions when its delegator loses theirs.
• An AI draft assessment stays a draft until an accountable person signs a separate assessment.

This is enforced by the Authority Kernel, not by prompt instructions. A model that is persuaded, manipulated or simply wrong still cannot cross the boundary between recommending and authorizing.

Every model endpoint is registered with the labels it may receive. Sensitive work is routed to models that run inside the customer boundary; a request whose data has no permitted route is refused rather than rerouted to a less restricted model. The model version is recorded with every output, so a finding can be traced to the model that produced it.

External and retrieved content is treated as data, never as instruction. It cannot alter policy, prompts or system behavior, and tools run only from a registry, within a declared side-effect class, in a sandbox without default network egress.

Governance constrains what AI may do; it does not make AI output correct. Evaluation bounds quality without guaranteeing it, and injected or poisoned content can still degrade an answer a person then relies on. Verity Command responds by keeping AI output attributable, cited and reviewable, and by keeping it on the recommending side of the chain. These limits are set out in the Public Threat Model.