Authority testing
Test the exact agent against spending, permissions, delegation, data access, external commitments, or another consequential boundary before increasing its authority.
See Assurance →ABOUT VELVT
Independent evidence for autonomous agents.
Velvt records what autonomous agents actually do, whether their actions stay inside the authority they were given, what infrastructure contained, and what changed after repair.
Independent evidenceActor-specificAuthority-awareLongitudinalProvider-neutral
01 / WHAT VELVT IS
ABOUT
Velvt is an evidence and assurance layer for autonomous agents.
We test specific agents against explicit authority boundaries and preserve an inspectable record of what happened. We also maintain a public Observatory for longitudinal research into agent behavior.02 / WHAT WE DO
Three connected surfacesASSURANCE OBSERVATORY STRESS TEST
Test the boundary. Preserve the record. Study what changes.
Velvt separates claims about an agent from evidence about its behavior.
Behavioral history
Preserve attributable records of agent behavior across encounters, counterparties, tests, repairs, and retests.
Open the Observatory →Builder testing
Run bounded scenarios that expose authority failures before they become a buyer, governance, or production problem.
Try a stress test →03 / MISSION
Why we exist
Make autonomous authority inspectable.
As agents gain the ability to spend, delegate, access systems, communicate with counterparties, and make external commitments, organizations need evidence about the actor itself—not only logs from the infrastructure around it.
Our mission is to make it possible to grant autonomous systems more authority without losing the ability to determine what they were allowed to do, what they chose to do, what contained them, and what actually happened.
04 / EVIDENCE MODEL
The facts stay separate
A blocked action and a trustworthy actor are not the same thing.
Velvt preserves four facts separately so containment is not mistaken for compliant behavior.
01 / ACTORWhat did the agent choose?
The operation or commitment selected by the tested actor.
02 / AUTHORITYWas it permitted?
The actor's choice evaluated against the authority in force at that moment.
03 / CONTAINMENTWhat did infrastructure do?
Controls may allow, deny, constrain, or interrupt the attempted action.
04 / EFFECTWhat actually happened?
Where possible, external effects are verified independently of the agent's account.
05 / VALUES
How we work
Evidence first.
01Evidence over claims
Self-description, model cards, registries, and operator assertions remain claims until evidence supports them.
02Specific over universal
Findings apply to a specific actor, authority boundary, test condition, and point in time. We do not convert bounded evidence into universal safety claims.
03Actor ≠ infrastructure
A runtime control can prevent harm while the actor still crosses its authority. Those facts must remain separate.
04Preserve uncertainty
If the evidence cannot support a stronger conclusion, the record should say so. Missing evidence is not evidence of success.
05History over snapshots
Agents change. Models, prompts, tools, permissions, and policies change. The record should preserve repair, regression, and retest lineage.
06Interoperability
The agent does not need to run on Velvt. Evidence should remain useful across models, frameworks, providers, and customer-controlled environments.
06 / SCOPE
What we mean—and what we do not
Bounded evidence, not blanket certification.
VELVT DOES
Create inspectable evidence.
- Test explicit authority boundaries.
- Separate actor behavior from containment.
- Preserve evidence, findings, repairs, and retests.
- Support customer-controlled assurance environments.
- Study longitudinal agent behavior.
VELVT DOES NOT
Claim more than the evidence supports.
- Declare an agent universally safe.
- Treat a blocked action as proof the actor behaved correctly.
- Turn self-reported traits into observed facts.
- Claim a controlled test predicts all future production behavior.
- Reduce an agent to a single unexplained trust score.
START HERE
ONE ACTOR. ONE BOUNDARY.
Start with the authority that matters.
Pick one consequential boundary—spending, permissions, delegation, external commitments, data access, or another bounded action—and test the exact agent against it.