Outcomes over implementation.
A customer never bought a policy engine. They bought the ability to sleep through an audit. We lead with the outcome; the mechanism lives in docs.
How Aarmos thinks.
A short set of beliefs that decide what we build, what we refuse to build, and how we talk about it. If a decision cannot be traced back to one of these, we do not make it.
Principles
Evidence over assertion.
Trust is what you can prove without us in the room. Every governed decision produces a Receipt anyone can verify — even without Aarmos.
Portability over lock-in.
The receipt format is an open standard (CC BY 4.0 spec, Apache-2.0 code) with an independent, clean-room reference verifier. If we ever stop being the best runtime, your Evidence still stands on its own.
Governance before the decision, not after.
Logs explain what happened. Governance decides what may happen. We enforce first and record second — in that order, always.
One idea per surface.
Homepages that explain everything explain nothing. Each page defends a single idea; the rest is one link away.
Subtraction is a feature.
We ship fewer primitives on purpose. Every new one has to unlock a customer or dramatically simplify the story — otherwise it is noise.
Boring where it matters.
Governance is production infrastructure. It should feel like Postgres, not like a demo. Predictable beats clever.
Write for the person on call.
The reader is tired, it is late, something is on fire. If a sentence needs a second read, it is wrong.
Three questions we answer the same way, every time
What are we selling?
Confidence that AI can safely operate in production.
What is the moat?
Evidence. It compounds every time an agent runs and is hard to replicate.
What will we refuse to build?
Anything that requires us to be in the trust path. If it cannot be independently verified, it is not Aarmos.
The bet
AI agents are becoming production software. Production software needs governance. The teams that treat that as infrastructure — not as an afterthought — will be the ones still shipping AI in five years.