Before intelligence can be trusted, it must learn to doubt.
AI systems retrieve, generate, and confidently hallucinate. When they are wrong, they do not notice, because nothing in how they work tracks the difference between what was observed and what was concluded. There is no mechanism for doubt, no architecture for uncertainty, no way to ask "why do I believe this?" and receive a real answer.
We believe aligned, trustworthy AI requires systems that think epistemically: systems that know what they know, why they believe it, and when to update. We publish openly because this problem is too important to solve alone.
Beyond Retrieval is the technical paper, introducing the Layered Epistemic Agent Protocol (LeAP) with stratified epistemic types, warrant functions, coherence invariants, AGM-compliant belief revision, and the Context In Tiered Epistemology (CITE) architecture with write-time coherence enforcement.
An Argument for Externalized Epistemics is the foundational argument, drawing on convergent evidence from information theory, optimization, interpretability, distributed systems, and neuroscience that epistemic state must live outside the models doing inference.
The reference implementation is Engrammic, and we welcome critique, extensions, and challenges. See CONTRIBUTING.md if you'd like to get involved.
CC BY-SA 4.0