Background
A national-lab consortium needed access to a foundation model for multi-scale materials work, quantum simulation, and a defense R&D programme. Their environment is air-gapped, ITAR-controlled and routinely audited. No commercial model API was an option.
The challenge
Most foundation models are exclusively cloud-based; their providers rely on telemetry and updates flowing back to a central environment. The consortium's requirement was the opposite: zero outbound traffic, model weights pinned to a version, and a refusal corpus that could be extended with the consortium's own scenarios.
What we did
- Aether deployed into an air-gapped environment with formal export-control certification.
- Model runs on internal hardware; weights are pinned per release; upgrade path is a physical media transfer with hash verification.
- Refusal corpus is extended with the consortium's own classified scenarios and red-team verified before each upgrade.
- Apex forward-deployed engineer cleared to the consortium's requirements; on-site for the install and the first three upgrades.
The outcome
- Zero data, zero prompts and zero telemetry cross the boundary.
- Three programmes now run on the deployment: a materials physics group, a quantum simulation group, and a classified R&D programme.
- Refusal corpus extension was completed in the first quarter; subsequent upgrades inherit it.
One detail worth calling out
The interesting engineering work was on the upgrade path. We had to demonstrate that a model upgrade does not change the refusal behaviour against the consortium's classified corpus — and that the proof itself does not leak corpus contents to us. Zero-knowledge proofs against a customer-controlled hash were the mechanism.