Glossary
The vocabulary, defined
Verillian is precise on purpose. These are the terms that recur across the platform and the documentation, for everyone on the buying committee, technical or not.
- Sentinel
- The lightweight agent on each workstation that intercepts AI traffic and enforces policy on the device.
- Admin server
- The central component for policy distribution, audit storage, fleet management, and reporting.
- Chain
- An append-only sequence of signed entries where each depends on the one before it, so any change is detectable.
- Chain entry
- A single signed record of one intercepted interaction, with timestamp, decision, and a hash linking it to the prior entry.
- Intercept
- Capturing AI traffic on the device using standard operating-system mechanisms, before it reaches a provider.
- Policy
- The institution's rules for what AI tools may do, evaluated at the tool and argument level at the moment of execution.
- Provider
- The downstream model service an AI tool reaches, such as Anthropic, OpenAI, or Google.
- Decision
- The outcome of a policy evaluation: allow, redact, or block. Every decision is signed into the chain.
- Fail-closed
- The rule that AI traffic stops when there is no valid policy or the audit pipeline fails. Unaudited AI is a stop condition.
- Opaque-ciphertext server
- A server that stores encrypted audit content it cannot read. Decryption requires the institution's key.
- Non-repudiation
- The property that an entry's origin cannot be denied, because it is signed by a device key the server cannot forge.
- Redaction
- Removing sensitive content from an outbound request before it reaches the provider.
Run a pilot on your own fleet
We work closely with our first institutions, on their endpoints, against their policies. The deployment becomes the proof.