ONE CONTRACT — A FOOTPRINT THAT GROWS WITH YOUR POSTURE
1
Validation profile. A containerized engine on your infrastructure — standard CPU servers, no external dependencies, in-memory restoration state. Enough to protect one real workflow end to end and measure detection on your own data.
2
Production target. A deployment-evolution step: the same engine with persistent, customer-controlled storage behind the vault interface, sized for your throughput. Placement follows one rule: detection, restoration state, and the restore path stay inside your controlled environment; providers receive tokens only.
3
Hardened production target. Key custody in your HSM/KMS, high availability, and isolated-network operation, added where the security posture requires them. Deployment evolution — the protection contract does not change.

What you stand up#

The current footprint is deliberately small:

  • Engine service — detection, tokenization, Context management, and restoration in one self-hosted service. Runs in a container on standard CPU servers; the standard detection stack has no GPU dependency.
  • Detection models — the self-hosted detection stack ships with the engine and runs entirely inside your environment. Nothing calls out.
  • Restoration state — the in-memory Context Store of the current profile, behind a replaceable storage interface; see Vault for how persistent deployments harden it.
  • Salus API — the integration surface your applications or gateway call. Text protection is the first lane; document pipelines follow, with Salus Desktop covering files on managed endpoints today.

What deploys where#

The engine deploys on-premises or in a private cloud tenancy you control. The placement rule is the same everywhere: everything that can re-identify a token stays inside your controlled environment. There is no phone-home, no license-server dependency, and no vendor access path into the deployment.

How a rollout runs#

  1. Validate. Run detection against your own data, in your environment, and read the measured result — not a datasheet number. One team, one workflow.
  2. Protect. Integrate the API in that workflow. Providers receive tokens; users keep getting useful answers. Failure behavior is agreed per data class before enforcement begins — the defaults fail closed.
  3. Expand. More workflows, more callers, more data classes; hardened storage and key custody come in when and where your posture requires them.

First Deployment walks this path in detail.