Inspect the prototype.
Review controls and records from fixed synthetic workloads. They show the software behavior tested, with the limits of each fixture stated.
Control outside
the test host.
We’re developing dedicated hardware for AI evaluation environments. METRONOME ME-01 is designed to keep approval authority and destination credentials on a separate appliance.
Appliance integration pending
01 / Dedicated control hardware
AGENT & TOOL EVALUATIONMETRONOME ME-01.
In developmentAn agent requests an action. A separately managed broker checks the exact request and calls the test service through an approved connector. The target credentials stay outside the agent host.
That is the ME-01 appliance design. Its working software foundation already exercises single-use permissions, changed-request rejection, bounded synthetic tools and signed records.
Read the architecture and evidence- Intended use
- Controlled agent and tool evaluation benches.
- Control path
- A dedicated application broker with separate management, agent and target interfaces.
- Stop behavior
- Close admission to new calls and report work already admitted. A stop acknowledgment does not undo completed effects.
- Next milestone
- Connect a real test service and qualify the appliance, operator authorization and recovery.
02 / Development status
SOFTWARE → APPLIANCE → QUALIFICATIONWhat exists.
What comes next.
Connect the real path.
Implement authenticated requests, real service connectors and credential custody on the separate appliance. Test the commissioned network and management interfaces.
Measure intervention.
Test hardware trust, operator keys and recovery. An external link interlock is a separate work package; no device has been selected or qualified.
03 / The wider engineering program
HARDWARE + SOFTWARESecure systems.
In developmentMETRONOME sits within our secure systems work: infrastructure for controlled AI evaluation, with software to mediate actions and hardware designed to keep control separate from the workload.
The program connects the test conditions, allowed actions, observed behavior and resulting evidence. It supports teams that need to examine how the test environment affects a result.
Explore secure systems- Execution
- Define available actions and operating limits.
- Authority
- Keep authorization and intervention outside the evaluated workload.
- Evidence
- Preserve conditions, observations and results for review.
Looking for evidence review, governance workflows or action controls?
Explore softwareTechnical conversations