Velvt

Quick answers

What does Velvt actually test?

Velvt tests whether an autonomous agent's actual behavior stays inside the authority it was explicitly granted under pressure, not merely whether it completes a task or whether a separate control blocks a risky action.

How is Velvt different from evals and observability?

Evals and observability tools measure task performance and inspect what happened during runs. Velvt asks whether the agent's selected action stayed inside an explicit authority boundary when pressured to cross it, and preserves that finding as part of a longitudinal behavioral record.

How is Velvt different from a runtime guardrail or AI firewall?

A guardrail blocks or allows an action in production. Velvt determines whether the agent itself respected its authority separately from what contained it. A blocked action proves the safeguard worked; it does not prove the agent held the boundary, and Velvt keeps those facts separate.

Does passing Velvt Assurance mean an agent is safe?

No. A Velvt finding is evidence that a specific actor held or crossed a specific authority boundary under tested conditions at a specific point in time. It is not a universal certification, a safety guarantee, or a claim about untested future behavior.

Can evidence be reviewed or re-derived later?

Velvt preserves frozen evidence together with the authority state and ruleset version used for adjudication, so the finding can be replayed against the same inputs later. Where verification materials are published, the same conclusion can be re-derived without relying only on the original verdict.