A Fragment of Kiafura on ARC-AGI-2: Frozen Runs, What They Transferred, and Why Static Memory Needs a Reasoner

Kiafura is a design for a machine that reasons without tokens and without text. A small part of it was built and run on ARC-AGI-2, in ARC Prize 2026, a benchmark whose tasks are grids, whose evidence is a handful of demonstrations, and whose contract allows two answers per test input. The fragment compiles verified solutions of a block of training tasks into frozen operational memory and applies it deterministically on an ordinary processor, with no training and no language model, behind a gate that emits an answer only when it reproduces every demonstration and otherwise names its abstention. This paper documents what was done with it, what it measured, and how the work continues. Each run compiled 100 tasks, was frozen, and was then measured on the 900 it had not seen, with a single answer: it solved between 39 and 63 further tasks, about forty further tasks for every hundred compiled, almost all of ARC-AGI-1 type, and 237 of its 255 answers on those tasks were exact. On the public evaluation set, five predictions were committed before the run: four were confirmed, the fifth was vacuously consistent, and no emission was wrong. Where several answers fit the demonstrations, something must choose, and the fragment has no organ for it. Every rule that chose in its place lost correct answers: requiring two supporting candidates removed five of the seven correct answers on fresh tasks, requiring all runs to agree abstained on ten tasks where one of two answers was right, and a majority of three runs to one was wrong. The two answers of the official contract keep such disagreements but do not resolve them: run by run, two answers solved exactly the tasks one answer solved. On ARC-AGI-2 the two official answers absorb every conflict observed, since no test input had more than two compatible answers; the zero score is a generation loss, answers never produced inside the compiled vocabulary. Choice binds when one answer is allowed or when compatible answers outnumber the slots. Choice among compatible answers cannot be frozen into memory; it needs information the demonstrations do not contain. That is the case, measured, for building the reasoner the design calls for, and Reasoning Theory says what it must bring.