Sovereign is the on-prem edition of GUARDIAN: same containment engine, hardened for air-gapped and regulated estates. No cloud, no telemetry, no external dependency — the agent's authority ends where your policy says it ends.
This is not a simulation. A real agent process is running on this server right now — every action you see was a real syscall gated by a real policy decision, and every audit entry below is a real sha256-chained record on disk. Watching is free for everyone; anything that touches the engine with your own data — custom intents, real process kill, external probes — requires a license.
Declarative boundaries per agent: files, processes, network endpoints, APIs, spend limits. Violations are denied before execution — not logged after.
Every decision written into a tamper-evident chain. Alter one entry and every hash after it breaks. Court-grade evidence of what the AI actually did.
Independent watchdog outside the agent process. Suspend, kill, or quarantine even if the agent is compromised or unresponsive.
Default-deny outbound. An agent cannot exfiltrate data to an external API because the connection is refused at the boundary.
Baseline drift on token usage, syscall rate, and data volume — flagged before it becomes an incident.
Nothing leaves the estate. No phone-home, no licence ping, no metrics. Sovereign works where the internet isn't allowed.
Single binary + policy file. No agents inside the agent — Sovereign supervises from the outside, so the agent doesn't need to cooperate.
# the policy enforcing the console above — verbatim: # /var/lib/sovereign/policy.json — ministry-finance-v3 agent: civis-docs-agent-04 fs.allow: [sandbox/docs/**, sandbox/output/**] fs.deny: [sandbox/citizens/**, /etc/**] net.allow: [127.0.0.1:8797, 127.0.0.1:443, 127.0.0.1:80] # loopback services only — no SSH, no lateral net.deny: [185.220.*] # known exit-node range → DENY proc.allow: [echo*, ocr-worker*] proc.deny: [/bin/sh, /bin/bash, curl, wget, nc]
And this deployment is live — press RUN, the response below is the real API answering right now:
/sovereign-api/state
— live decisions + chain tip
/sovereign-api/verify
— re-verify the whole ledger
/sovereign-api/report
— compliance report as text
Static binary, ~9 MB. Linux x86/ARM, Windows, container sidecar, or VM image.
On-prem, private cloud, air-gapped. Policy signed and versioned like code.
One operator console governs many agents across hosts — policies pushed, ledgers aggregated locally.
| Requirement | What it demands | Sovereign |
|---|---|---|
| EU AI Act — Art. 12 | Automatic logging over the system's lifetime | Hash-chained audit ledger |
| EU AI Act — Art. 14 | Effective human oversight, incl. intervention & stop | Kill switch + suspend/resume |
| ISO/IEC 42001 | Operational controls for AI management systems | Policy engine + audit export |
| NIST AI RMF | Monitor & govern AI behaviour in operation | Runtime sensors, deny-by-default |
| Data residency | No foreign telemetry, on-prem control | Air-gap capable, zero phone-home |
Recomputes the entire sha256 chain of the audit ledger on the server, right now — not a screenshot. If any entry were altered after the fact, this check fails.
The containment engine for builders and teams — local supervision, kill switch, audit log. guardian.cam →
The on-prem edition: signed policies, fleet console, compliance exports, air-gap operation. For deployments where governance is a contractual requirement.