I’ve spent the last year building SAFi, an open-source cognitive architecture that wraps around AI models (GPT, Claude, etc.) to enforce alignment with human values. Safi is a "System 2" architecture inspired by classical philosophy. It separates the generation from the decision:

The Intellect: proposes a draft.

The Will: decides to block or approve the drafts.

The Conscience: audits the drafts based on set core values

The Spirit: An EMA (Exponential Moving Average) vector that tracks "Ethical Drift" over time and injects course-correction into the context window.

Feedback is welcome!