02 · Principles 03 / 12

How we build with care

Safety as bedrock, not checklist

Every model undergoes six months of adversarial evaluation before any public release.

Alignment is a long game

We fund interpretability research that may not return results this decade — and that is the point.

Transparency by default

Reasoning traces are visible to every user, not gated behind an enterprise API tier.

Restraint as competitive advantage

We have declined launch requests when evaluations show unresolved risk — and we would do it again.