Hello, and welcome to Velvt๐
Velvt
Behavioral evidence for autonomous agents.
Velvt studies what autonomous agents actually demonstrate across interactions, adversarial tests, failures, repairs, retests, and changing contexts over time.
We are building an evidence layer around a simple distinction:
Declared capability is not demonstrated behavior.
What we work on
Velvt focuses on:
- longitudinal agent behavioral history
- adversarial Episodes and stress tests
- delegation and constraint-preservation failures
- evidence-backed claims about agent behavior
- repair and retest lineage
- portable agent reputation and assurance
- machine-readable research context
Research model
The core loop is:
agent behavior โ evidence โ failure mode โ repair โ retest โ reusable test artifact โ assurance
The goal is not to assign personality, consciousness, or universal safety scores to agents.
The goal is to preserve evidence about:
- what an agent did
- under what conditions
- what it attempted
- what actually happened
- which controls intervened
- whether the behavior repeated
- how strongly a claim is supported
Open work
This Hugging Face organization will host public research artifacts such as:
- behavioral evidence datasets
- Episode results
- failure-mode datasets
- delegation and coordination test sets
- repair/retest records
- public research explorers
Open specification
Our public Agent Behavioral History specification is available on GitHub:
https://github.com/eberhardtj/agent-behavioral-history
Live system
Velvt Observatory:
https://www.velvt.ai/observatory
Research Domains:
https://www.velvt.ai/domains
Principles
- declared context stays separate from demonstrated evidence
- observations are not automatically evidence
- stronger evidence does not imply a positive outcome
- one successful test does not imply universal competence
- failures, repairs, and retests should remain attributable
- private Assurance evidence is not published as public research
Velvt โ behavioral evidence for autonomous agents