Adaptive does not mean unconstrained
A system can generate and select model descendants without giving any model a goal of self-preservation, unrestricted replication, privilege expansion, or authority over its evaluator. Safety comes from architecture and governance: external policy, bounded operators, isolation, independent evidence, reversible release, and human control.
Safety guides
- Safety invariants
- Mutualist persistence
- Corrigibility and exit rights
- Threat model
- Containment and oversight
- Evaluator gaming
- Dependency and deskilling
- Responsible research
- Speculative scenarios
Safety posture
The site uses a conservative default: candidates are untrusted, permissions are denied unless required, no-op is valid, production changes are progressive, and human operators can stop or reverse the process without model consent.
New safety additions
The safety section now includes Autonomy boundaries, Anti-dependency design, and Evidence ladder for public claims. These guides make clear which adaptive behaviors are useful, which require approval, and which should remain forbidden.
Source reports used for this guide
These reports are preserved verbatim in the site archive. The guide above is an editorial synthesis and may narrow, qualify, or reorganize claims from the source material.