Vervain Govern what an agent thinks before it acts.
Inference Governance
Vervain Govern what an agent thinks before it acts.
Inference Governance
Your identity and authorization systems decide whether an agent may read a customer record. They do not show where that record went next.
One agent turn can leave copies in a context window, a reused KV cache prefix, a vector index, a tool result, a log line, and a checkpoint. Each has its own lifetime and its own owner. None of them report back to the system that approved the read.
So when a customer asks you to delete their data, or someone questions a tenant boundary, the honest answer today is an estimate. The gap widens with every agent you deploy.
Vervain traces a specific record forward, from the moment an agent reads it to every place a derivative came to rest, and reports what a revocation reached and what it did not cover.
It runs inside your environment, on vLLM, SGLang, and Ollama, down to the KV cache the runtime reuses between requests. It reports what it observed and states the limits of that observation in the same report. A surface it cannot instrument is named rather than assumed clean.
A one-week cross-tenant contamination assessment. Can you show today that one tenant's data has never appeared in another tenant's context? We are taking a small number of design partners now.
We are also building an open-weight model optimization product. Open weights are the only place you can govern and optimize the same inference path, and we intend to own that path. It runs on vLLM, SGLang, and Ollama. We are opening early access now. If you self-host open-weight models, write to us.
- Vervain Team