Cordon.
Exfiltration is architecturally impossible.
Cordon is the sovereign foundation every other Regnant system runs on. No data, query, weight, or result ever crosses the institution's perimeter. Not a privacy policy, a set of technical constraints enforced from a hardware root of trust to a client-verifiable, Merkle-chained audit log.
Four layers, each enforcing the one above it
- L4Offline-Verifiable Audit Log
- L3Encrypted Model Weights
- L2Trusted Execution Environment
- L1Hardware Root of Trust
Identification
- Designation
- RGN-01 · Cordon
- District
- Sovereign Infrastructure
- Function
- Private Inference Engine
- Source
- Open
Answers
“Our AI cannot touch the public cloud. What do we run?”
- TPM 2.0 attestation and Secure Boot: containment enforced in hardware, not policy
- AMD SEV-SNP, Intel SGX v2, or ARM TrustZone: inference runs in memory no operator can inspect
- AES-256-GCM weights under a key hierarchy the institution alone holds. The vendor cannot decrypt
- Ed25519-signed Merkle audit chain, verifiable by the client with no internet connection
Architecture
Scroll down to bedrock.
Four layers, each one enforcing the one above it. The page you are reading descends the same way an auditor would: from the log they can verify, down to the silicon that makes it true.
Offline-Verifiable Audit Log
- Ed25519-signed Merkle chain
- Every query, response, and state transition recorded
- Client verifies the entire chain without internet access
The layer the auditor sees. Everything below exists so this one can be trusted.
Encrypted Model Weights
- AES-256-GCM symmetric encryption
- Client-held key hierarchy: the vendor cannot decrypt
- Per-deployment key rotation with zero-knowledge proof
Provisioned once, owned forever. We hand over weights we can never read back.
Trusted Execution Environment
- AMD SEV-SNP encrypted memory isolation
- Intel SGX v2 enclave execution, where available
- ARM TrustZone secure-world partitioning
Inference runs inside memory no other process: or operator, can inspect.
Hardware Root of Trust
- TPM 2.0 cryptographic attestation
- Secure Boot with verified firmware chain
- dm-verity read-only root filesystem validation
Bedrock. Containment is enforced here: in hardware, not policy.
The audit log, live
Don't trust the audit log. Check it.
"Offline-verifiable" is a provable claim, so this page proves it: a live Merkle chain, hashed and re-checked by your own machine, no internet required.
SHA-256 via your browser's WebCrypto · in production every link is also Ed25519-signed
- 0001QUERYinference request received…
- 0002RESPONSEinference result returned…
- 0003STATEmodel weights attested (TPM 2.0)…
- 0004QUERYinference request received…
- 0005APPROVALhuman decision recorded…
- 0006STATEkey rotation completed…
The chain above was hashed live when this page loaded. Try to catch it lying.
Who runs it
01
Central banks
Sovereign monetary policy intelligence with zero exfiltration risk.
02
Defence intelligence
Classification-aware analysis that never leaves the wire.
03
Critical infrastructure
Decision systems for energy, water, and telecommunications.
On-premise handover
What the institution receives.
They own the roadmap. We hand over the system: not a rental.
- 01Encrypted model weights: provisioned once, owned forever
- 02An open, auditable inference runtime
- 03A key hierarchy the institution alone holds
- 04An air-gap option for fully offline operation
- 05Update timing, testing, verification, and approval, the institution owns the roadmap
Built on this ground
Next system
02 / 08