Kiafura: An Externalized Reasoning Architecture

Kiafura is a design for a reasoning machine that reasons without tokens and without text. It was conceived to overcome the limits of a field that grows by making one model ever larger and ever more energy-hungry, and to put an alternative to the test. Its reasoning state is a typed causal graph; at run time language models only translate between the machine and the world and have no right to reason; at construction time they act as engineers and, by Reasoning Theory's attribution rule, belong to the reasoner until its memory is frozen, as in the built fragment (Fiorilla, 2026c). Its memory is designed to be deterministic and to grow without limit, with no context window: what the machine knows and the project it is working on are kept as raw, immutable knowledge and reorganised continuously, and the reasoner reads only the small intersection of the two that the current step needs. Every claim leaves through a gate that can abstain, so the machine makes no unsupported claim: it states what it has verified, or says what it cannot settle. Because memory and reasoning are delivered on demand, the active reasoner is designed to be small, and ideally to run on ordinary processors (Hypothesis 6.1). As a hypothesis, Kiafura aims at reasoning on a par with today's best systems, to the extent that the datasets compiled into its static memory can be made deep and well organised, at a small fraction of their energy (Hypotheses 8.1 and 15.1); and it is designed to learn from its own verified experience instead of waiting to be retrained. Several of its organs answer functions that Reasoning Theory shows no reasoner can drop. A small part of it was built and measured on ARC-AGI-2, in ARC Prize 2026, a testbed with no text to reason with.