Anchor Manifesto
Capability without governance is not progress.
Anchor is founded on a simple premise: increasingly capable systems require increasingly reliable mechanisms of governance.
The Current Assumption
Much of modern AI safety relies on probabilistic compliance.
Systems are trained to behave.
Systems are encouraged to follow rules.
Systems are evaluated against benchmarks.
Yet the underlying assumption remains: the system chooses to comply.
The Problem
Institutions do not operate on trust alone.
Financial systems use controls.
Operating systems use permissions.
Aviation relies on procedures, verification, and containment.
Yet many intelligent systems are expected to remain safe through instruction and behavioral optimization alone.
Anchor questions whether that approach can scale.
The Anchor Thesis
Governance should not depend entirely on model behavior.
Governance should exist as an independent layer.
Constraints should be enforced.
Capabilities should be isolated.
Actions should be auditable.
Policies should remain active during execution.
What Anchor Explores
Long-Term Vision
Anchor is not an attempt to build more capable systems.
Anchor is an attempt to understand how capable systems can remain governable.
The future of intelligence may depend as much on governance as capability.