The Core Ring: Composition, Not Components, Is What Large Language Models Lack

Abstract

Large language models have improved chiefly by scaling one monolithic computation. We argue the binding constraint is no longer missing components—language, reasoning, and knowledge are all present—but a missing composition in which they form a closed loop. We propose the Core Ring, an architecture with no single core. Four functional conditions are claimed to be necessary for cognitive agency—semantic intake, distribution, epistemic self-modification, and self-assessment. A reference implementation realizes them as four mixture-of-experts (MoE) modules: a Linguistic MoE normalizing input into a language-independent semantic skeleton, a Scheduler MoE distributing computation, a Learning MoE acquiring and consolidating knowledge, and a Metacognitive MoE arbitrating by fixed rule. These conditions are necessary, not sufficient; goal systems, world models, and embodiment are out of scope. A fifth Executor MoE separates alive from usable. Experience changes not only what the system knows but what it computes. We further argue that Engrams and MoEs are one substrate in two regimes joined by an explicit promotion rule; that metacognition must hold final disposition, must not vote, and must compute relevance internally; and that forgetting is declining accessibility, not lost memory. This is a position paper: it reports no experiments, and every claim needs validation.