The Golem, Not the Genie: A Framework for Transient Cognitive Architectures in Large Language Models

We argue that frontier LLMs are assembled automata, not hidden minds. A prompt functions as a blueprint that configures a temporary Transient Cognitive Architecture (TCA) from pre-learned parts; successes and failures then reflect the architecture assembled at inference, not motives. This framework addresses the field's ``missing middle''---providing an Algorithmic Level account Marr1982Vision that bridges the gap between low-level circuits and high-level behavior. We formalize this via the Analogy Engine, a meso-level mechanism that maps a prompt’s relational structure onto reusable circuits. We also outline ACC, a candidate micro-level realization of the Analogy Engine, while keeping the core account implementation-neutral. Using this lens, we re-explain established phenomena (e. g. , agreement under conflict, scratchpad→answer divergence, transfer of adversarial “anchors”) as architectural reconfigurations or collapses, not choices. This is a theory-first paper proposing preregistered experimental protocols rather than reporting new empirical results. Our contribution is (i) a precise, mechanism-level account that unifies prior results; (ii) a suite of discriminative predictions testing for architectural state-shifts, functional boundaries in reasoning, and a core preference for structural patterns over surface cues. If borne out, this theory reframes alignment and reliability as engineering problems of assembly, routing, and constraints. This yields immediate practical levers—such as blueprint hygiene, phase isolation, and collapse detection—and offers cheap, decisive tests for validation. Even null results would remain informative by narrowing the scope of architectural claims.