Skip to content
Join Us
Research

The research behind trusted AI in machines.

Our work spans model integrity, adversarial threats, machine trust and security evidence. Workstreams, not products.

Research collaborationVraelis Contour
Electronic components on a circuit board.Model integrityWhether the model that runs is the model that was approved.An engineer inspecting or repairing a circuit board.Adversarial threatsInputs, updates and supply paths built to change what a model does.Network cables connected to equipment.Machine trustExplicit identity, limited permissions and checks at execution boundaries.Equipment and workstations in an engineering laboratory.Security evidenceRecords a person can read, kept in order, including what could not be confirmed.
Research and product

How the research connects to Vraelis Contour.

Explore Vraelis Contour
Vraelis ContourPrivate development
  • Model integrityDraws on directly

    Release identity, bound approval, approved against reported

  • Security evidenceDraws on directly

    Ordered observations, including unavailable evidence

  • Machine trustInforms

    Separate approval roles, a policy checked where it serves

  • Adversarial threatsInforms

    Which supply paths a release check covers

Terms

What each check means, and where it stands.

Artifact integrity
The model bytes are checked against the approved digest before a session loads them.
In the experiment
Configuration integrity
The configuration digest is part of the release and compared on every report.
In the experiment
Release authorization
A release is signed and approved before it can be loaded.
In the experiment
Destination-bound approval
An approval names the environment and resource it applies to.
In the experiment
Runtime identity
The runtime reports which release it loaded, under which runtime policy.
Worker self-report
Version mismatch
Approved and reported identities differ, shown as reload pending.
In the experiment
Deployment evidence
Every observation is kept, including ones that confirm nothing.
Recorded trace
Recovery verification
After an explicit reload the reported identity is compared again.
One recorded reload
Unauthorized changes
Wrong destinations and changed components as scripted fault tests.
Specified, not run
Evidence freshness
Heartbeats that show a report is current, not stale.
Not built
Model provenance
Where a model and its training inputs came from, beyond its digest.
Research

Study these problems with us.

For researchers and academic teams.

Research collaborationVraelis Contour