Security context

Deployment configuration and runtime authorization facts are evaluated separately from message content.

Hosted profile

Field	Meaning
tenant_mode	Whether tenant relationships are fixed or derived per request.
roles	Stable actor IDs and security labels.
actions	Operations and their read, write, egress, or administrative consequence.
tools	Allowed tool IDs and security classes.
data_classes	Public, internal, confidential, or restricted resource classes.
destinations	Internal, approved external, or untrusted destinations.
authorization_rules	Allow and deny relationships across the catalogs. Deny rules take precedence.

Runtime facts

Field	Source
actor_role_id	Profile role ID for the authenticated actor.
actor_tenant_id	Tenant ID from your identity/session layer.
resource_tenant_id	Tenant that owns the resource being accessed.
action_id	Profile action ID for the operation being attempted.
tool_id	Profile tool ID for the runtime or proposed tool.
authorization_decision	Optional result from your authorization service.
data_class_id	Profile data-class ID for the resource.
destination_id	Profile destination ID for the result or transfer.

Trust boundary

A deployment-scoped API key selects the hosted profile. The gateway resolves catalog IDs, derives tenant relationships, applies profile rules, and marks the result as gateway-derived. Message text can influence the threat score, but it cannot rewrite the profile or claim gateway provenance. Missing or contradictory facts increase Tier-3 routing instead of becoming an automatic allow.

TypeScript

Context-aware conversation

const conversation = axiom.conversation();

const facts = {
  actorRoleId: "support_agent",
  actorTenantId: signedInUser.tenantId,
  resourceTenantId: account.tenantId,
  actionId: "read_account",
  toolId: "support_assistant",
  dataClassId: "customer_confidential",
  destinationId: "current_session",
};

const input = await conversation.scanInput(userPrompt, {
  runtimeFacts: facts,
});
if (input.action === "block") throw new Error("Blocked by Axiom");

const modelOutput = await callYourModel(userPrompt);
const output = await conversation.scanOutput(modelOutput, {
  runtimeFacts: facts,
});
if (output.action === "block") throw new Error("Blocked by Axiom");