01 / Controlled execution
Define the conditions.
Set the workload’s scope, available actions, and operating limits. Evaluate behavior in an environment whose boundaries can be examined.
A boundary outside
the AI workload.
We’re developing the infrastructure for controlled AI evaluation: software to mediate actions and preserve evidence, with dedicated hardware designed to keep control separate from the system under test.
Discuss a technical walkthroughDedicated control hardware / In development
CUTPLANE.
Authority outside the workload.
CP-01 is the first appliance design in the program. Explore exact-request authorization, the independent containment architecture, and a verified playback of an archived software test. CP-01 hardware integration is the next development stage.
Explore CUTPLANEControl. Evaluation. Evidence.
Conceptual architecture02 / ControlSeparate boundaryAuthorization & intervention
03 / EvaluationAI workloadDefined test environment
04 / EvidenceRecord the conditions. Examine the result. Test again.
What we’re engineering
The environment is
part of the evaluation.
02 / Independent control
Keep authority separate.
Develop a dedicated control appliance and supporting software for authorization and intervention outside the evaluated workload.
03 / Inspectable evidence
Preserve what happened.
Connect conditions, observed behavior, and test results in a verifiable record. Keep the basis for the next decision available for review.
Development status
A working foundation.
A defined engineering path.
From tested software controls to dedicated hardware, the program connects authorization, evaluation, and evidence.
Software prototype
- Executable evaluation and control prototypes using fixed synthetic workloads.
- Tests of authorization, bounded execution, intervention, and evidence integrity.
- Recorded examples that support inspection of the software behavior tested.
Dedicated hardware
- Architecture and integration work for a separate control appliance.
- A construction and commissioning plan for physical qualification.
- Qualification work covering hardware trust, physical intervention, and recovery.
Where the work is directed
When the test environment
needs its own evidence.
Frontier evaluation
Controlled research environments for sensitive systems and difficult evaluation questions.
Agentic workflows
Examine how authorization and operating boundaries affect systems that can take actions.
Institutional assurance
Support reviewers who need to understand how the environment shapes a result and its limits.
Engineering conversations
Bring the system.
Define the boundary.
Explore the prototype work with our team. Bring a use case, a research question, or a capability to develop together.