What is ControlCore?
ControlCore is the human-authority control plane of AI Control. It evaluates every autonomous request through identity verification, capability check, policy evaluation, context evaluation, risk evaluation and — when required — human approval, returning an enforceable decision (allow, limit, quarantine or block) that execution gateways act on and the audit trail records.
- Default deny when identity, capability or policy is missing.
- Most-restrictive matching policy effect wins.
- High-impact actions always require human approval.
- Requesters cannot approve their own requests.
How it works
- Identity verification — The caller presents an attested identity. Unknown or revoked identities stop here with BLOCK.
- Capability check — The requested capability must be explicitly granted on the target resource and environment.
- Policy + context — Structured policies are evaluated against the request and its context — environment, target, time, workspace.
- Risk evaluation — Weighted factors produce an explainable score that policy thresholds act on.
- Human approval — Production writes, robot motion, physical-world actions and fund movements are held for an accountable human.
- Decision, execution, audit — Gateways execute only scoped, short-lived grants; every step is written to the hash-chained audit log.
Frequently asked questions
What happens if an AI agent is not registered?
It is blocked. AI has no authority by default, so an unknown identity cannot pass identity verification.
Can a policy make a decision less strict than the built-in guardrails?
No. Policies can only make a decision stricter. High-impact actions always require human approval regardless of policy.
What is the difference between LIMIT and REQUIRE_APPROVAL?
LIMIT allows the action within reduced scope (for example read-only or a record cap). REQUIRE_APPROVAL holds the action until an authorized human approves it.
Is every decision audited?
Yes. Inputs, matched policies, risk factors and outcome are recorded in a tamper-evident hash chain that can be verified with GET /api/v1/audit/verify.