INTELLIGENCE
IS NOT
AUTHORITY.
The future can become more capable
without becoming less ours.
A safety promise is only as strong
as the power to enforce it.
Binding release gates. Protected independent review. Funding that survives a pause. Human authority over both the machine and the institution.
If your safety case requires
the system to cooperate,
keep working.
Assume a consequential AI system may deceive, seek power, and resist correction. Then ask what must remain true even if it tries.
We are forming a lab around that obligation: bounded authority, adversarial evidence, and an institution with the power to stop.
What the lab must establish ↗Who can say no
when it matters?
Follow the objective. Follow the ownership. Find where human authority becomes optional.
↗02 / ESTABLISH THE REQUIREMENTSMake the limits
survive resistance.
Define the prohibited outcome, build the barrier, and let an adversary try to break it.
↗03 / FORM THE LABBuild the team
the claim requires.
Six research functions. Concrete first deliverables. Independent authority to halt expansion.
↗The machine must not write
its own permission slip.