Security by construction.

PRINCIPLES

• AI does not decide what it is allowed to do. Policy does.
• Untrusted input lowers trust — and policy acts on trust at the moment of action, not at the moment of input.
• Autonomy does not mean absence of control — high-risk actions stop for a human.
• Who watches the control plane? Sentinel does, from outside its trust domain.
• Keep the evidence — immutable, sealed, verifiable.
• Own the infrastructure — no vendor control plane in the path.

SCENARIO — DATA EXFILTRATION

An AI agent attempts to send confidential information outside the company. User request → agent starts → internal data accessed → external web result added → trust drops → email action requested → Imperium policy blocks → Sentinel records integrity state → Observatory records trace → evidence package created. ACTION PREVENTED BEFORE EXECUTION.

SCENARIO — UNAPPROVED MODEL

An employee uses an unapproved model. Discovery identifies the client and endpoint; policy classifies it unmanaged; routing offers the governed alternative; the owner is notified; the evidence shows what, if anything, was sent.

SCENARIO — COST SPIKE

AI costs suddenly spike. Cost attribution finds the agent and session; routing policy moves the workload to a local model; a human is asked before the budget is exceeded.