# Defense

Model release integrity for defense systems. Vraelis does not build weapons or mission software.

> **Intended, not deployed.** Intended application areas. Vraelis has no customer deployments in them.

## A model update crosses organizations.

### Updates cross organizations

Trained by one team, integrated by another, operated by a third.

### An unexpected change is a security and operational risk

A different release, or the wrong destination, changes what the system does.

### Reviewers need to read the evidence

What was approved, for where, and what the system reports running.

## The workflow, in this area

1. Identify the release.
2. Bind approval to the release and destination. Destination-bound approval: test approval cannot carry into operations.
3. Integrate with a supported runtime.
4. Compare approved and reported.
5. Preserve discrepancies, missing evidence and recovery.

![Approval for cpu-2: destination, policy, a 60 second age limit and the runtime.](https://vraelis.com/v8/contour/crop-release-approval.webp)

*Read-only evidence app, inside the controlled ONNX reference experiment.*

## The builder should not be the only judge.

- Security software only. Vraelis does not decide how a system is used.
- Approval apart from the builder. In the experiment, release authorization separates publisher, deployer and approver roles. The trace keeps a receipt digest, not the approver's identity.
- A record kept in order. Every observation is kept, including ones that confirm nothing.

## Not claimed

- No deployments in defense systems, and no customers.
- No contracts, security clearances or government authorizations.
- No access to classified or restricted environments.
- No certification, and no claim of fitness for operational use.
- No protection against adversarial inputs or perception attacks.
- No hardware attestation. Runtime identity is a worker self-report.

- [Defense on vraelis.com](https://vraelis.com/applications/defense): The page this summary comes from.
- [Model integrity](https://docs.vraelis.com/model-integrity): Related research.
- [Machine trust](https://docs.vraelis.com/machine-trust): Related research.
- [Adversarial threats](https://docs.vraelis.com/adversarial-threats): Related research.

---

Vraelis documentation, Vraelis. https://docs.vraelis.com/defense