This paper specifies VDR-LLM-Prolog, a language model architecture where every value is an exact fraction, every derivation is recorded in a logic programming knowledge base, every constraint is a first-class queryable object, and every piece of knowledge is directly surfaceable to the user without passing through the language model's token generation. The specification integrates four prior results: VDR exact arithmetic (VDR-1 through VDR-3), the VDR machine learning stack (VDR-4), transcendental constant representation (MATH-3/MATH-4), and a custom Prolog-style knowledge engine designed for LLM provenance. The system has three layers. The arithmetic layer (VDR) ensures every number is an exact fraction with zero drift and zero silent truncation. The logic layer (Prolog) records how every value was derived, what it depends on, and what constraints it satisfies. The conversation layer manages scoped knowledge bases, working data sets, topic tracking, and constraint activation — giving the language model structured persistent memory that survives topic switches, supports inheritance and shadowing, and is directly queryable by the user. The central claim is that data provenance, constraint enforcement, and conversational state tracking are not features to be bolted onto a language model after the fact. They are architectural requirements that should be present from the foundation. This paper specifies what that foundation looks like.
Geoffrey Howland (Fri,) studied this question.