How we build,
what we check,
what we show.
Centra is an agent platform built by people who wanted an answer to a simple question: "How do we know what it did?" This hub covers how we evaluate safety, govern agent behavior, and report what the system actually did.
Two documents. One thesis.
Safety & Evaluation Report
How we test Gideon before it touches real work. Internal testing, external review, safeguard architecture, and what happens when something goes wrong.
Responsible Practices
Our framework for risk assessment, agent governance, security, privacy, and the societal commitments we hold ourselves to as we scale.
What we hold ourselves to.
Humans decide. Machines execute.
Every consequential action (spending money, writing to external systems, anything that cannot be undone) stops and waits for a human. This is not a setting. It is architecture.
Every action is journaled.
Inspect recorded activity, tool results and saved files to understand what happened during a task.
Receipts you can keep.
Every finished run ends in a signed receipt: inputs, outputs, cost, hash. Proof of work as an artifact, not a picture of one.
Absence is visible on purpose.
Models cannot see what is not there. Centra makes missing steps, unverified facts, and incomplete tasks visible, so they can be corrected instead of ignored.
Security is a phase.
Before a single line of code is planned, the security specialist attacks the design. Nothing proceeds until it has been hardened.
We publish how we work.
This transparency hub, our methodology document, our legal framework, and our product documentation are written to be read.
We report safeguards against the product that actually exists. A control observed in one environment is not automatically claimed for another, and a research design is never presented as a qualified release.
We will update these pages as our practices evolve. When we add new safeguards, change our evaluation methodology, or reach new milestones, this is where you will find it. The version label on each page tells you when it was last reviewed.