Author here. I'm excited to share the first public prototype of CORE, a project I've been working on to explore how an AI system can build and evolve software safely.
I've always been concerned by the "black box" nature of many AI dev tools. They generate code, but often without clear principles, governance, or traceability. This makes them risky for anything that requires trust and auditability.
CORE is my attempt at a solution. The core idea is a strict separation between the system's "Mind" (a declarative .intent/ directory with its constitution, policies, and self-knowledge) and its "Body" (the Python source code). It can only act in ways that are aligned with its constitution, and any proposed changes to its own governing principles must go through a secure, human-in-the-loop "canary check" process.
Right now, it's an architectural prototype. The self-audit and constitutional amendment loops are stable and working. The next big step is building out the agent capabilities to let it generate new applications from scratch.
I'm making this public now to get feedback on the core architecture and to find potential collaborators who are also interested in this new frontier of governed, self-improving systems.
I'd love to hear your thoughts. I'll be here all day to answer questions.
Author here. I'm excited to share the first public prototype of CORE, a project I've been working on to explore how an AI system can build and evolve software safely. I've always been concerned by the "black box" nature of many AI dev tools. They generate code, but often without clear principles, governance, or traceability. This makes them risky for anything that requires trust and auditability. CORE is my attempt at a solution. The core idea is a strict separation between the system's "Mind" (a declarative .intent/ directory with its constitution, policies, and self-knowledge) and its "Body" (the Python source code). It can only act in ways that are aligned with its constitution, and any proposed changes to its own governing principles must go through a secure, human-in-the-loop "canary check" process. Right now, it's an architectural prototype. The self-audit and constitutional amendment loops are stable and working. The next big step is building out the agent capabilities to let it generate new applications from scratch. I'm making this public now to get feedback on the core architecture and to find potential collaborators who are also interested in this new frontier of governed, self-improving systems. I'd love to hear your thoughts. I'll be here all day to answer questions.