Seven Layer
GPU Model Protection

Sovereign mode

Run everything on your own hardware, under your own key — no control plane, nothing leaving your perimeter.

For a bank, a regulator, or any workload where the answer to "who has access to the box" must be nobody but you, run in sovereign mode. SCSA is a set of binaries that run on your GPU host. There is no control plane in the middle and no external party authorized on the machine.

What runs on the node

BinaryWhat it does
scsa-vaultSeal and unseal to the attested GPU; attest; anti-replay and Confidential-Computing self-tests
scsa-sealmodelSeal a safetensors model, and unseal it
scsa-inferUnseal weights into VRAM only and run inference; prove in-VRAM consumption
zk-cliGenerate and verify transparent STARK audit proofs
scsa-verifyCheck a decision envelope offline, from bytes alone

They are self-contained and build from source on your hardware. Only a CUDA 12.x driver is required — no nvcc, no repo checkout, no external download. A sovereign deployment can run fully offline.

Sealing a model

scsa-sealmodel seal model.safetensors -o model.scsa --model-id my-model
scsa-sealmodel unseal model.scsa -o model.safetensors

The key is derived from your GPU's identity and never leaves the machine. The .scsa is inert on any other hardware. scsa-vault check model.scsa verifies the GPU matches without decrypting.

Running inference without the weights ever landing

scsa-infer run model.scsa --input <data>
scsa-infer vram-proof

Weights are decrypted in GPU memory only. They never exist as files. vram-proof demonstrates that the decrypt-and-consume path stays inside VRAM.

Recovery, deliberately chosen

scsa-vault seal secret.bin -o secret.scsa --shards 3 --threshold 2
scsa-vault seal secret.bin -o secret.scsa --single-device

The default is 2-of-3 recovery shards, because binding to one GPU means GPU loss is data loss, and most people do not want that discovered during an incident. --single-device opts out and is exactly as absolute as it sounds — there is no recovery.

--expiry <HOURS> and --max-unseal <N> bound the seal in time and in uses.

Attestation tiers

scsa-vault reports which tier it can actually reach on your hardware, and refuses rather than pretending:

TierBinding
1GPU measurement + driver/VBIOS, no TPM
2GPU measurement + TPM-sealed secret, bare metal
3GPU fused key + TPM in Confidential Computing mode — not yet available
scsa-vault attest --nonce <64-hex>
scsa-vault verify-replay      # an attestation for one challenge must fail against another
scsa-vault cc-check           # fail-closed Confidential Computing enforcement

On a GPU without Confidential Computing, attest and cc-check fail closed — the gate correctly refuses rather than pretending. See Proofs for how to read the verdicts.

Why this is the sovereign path

  • No control plane. The binaries run on your box. There is no server to trust, meter, or compromise.
  • Your key only. Nothing from the operator is authorized on the machine.
  • Air-gap friendly. The binaries phone home to nothing.

The strongest posture

Sovereign mode removes the operator from the loop. To also remove the need to trust any single machine or party — including your own operators — combine it with threshold custody, where the key is split so nobody holds enough to unseal alone, and with the Authority Fabric if you want each use to be a governed, logged, offline-checkable decision.

Security then rests on the math: hardware-signed attestation plus threshold cryptography.

On this page