A company structured so that the mission is not the first thing dropped under pressure.

Two jobs that keep interfering with each other

This company exists because a group of researchers concluded that the safety question would be decided by whoever was actually building frontier systems, and that abstaining from the work would not protect anyone. So we do both: develop the models, and run the research programme that makes their behaviour legible.

The arrangement is deliberately awkward. The people measuring risk report into the same organisation that benefits from shipping, and the escalation path runs to a board committee rather than to a launch team. We think that tension is a feature; a safety function that never inconveniences anyone is not doing anything.

What we hold ourselves to

• Publish the result you did not want. Negative findings and corrections go in the same feed as releases.
• Tie commitments to measurements. A promise without a threshold and a sign-off is a preference.
• Keep a human accountable by name. Every tier decision has an owner who can be asked about it.
• Prefer being corrected early to being trusted wrongly. External review, including of our methods.
• Build for the person who has to debug this in three years. Legibility over cleverness, internally too.