The Governable Layer: A Constructive Corollary to Why Stochastic-AI Regulation Is a Fallacy

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

A prior paper argued that the planetary debate over stochastic-AI regulation is a fallacy inthe precise sense: not a mistake of degree, but a category error of kind. Why Stochastic-AI Regulation Is a Fallacy—And Yet, Planetarily Discussed." Zenodo, 2026.DOI:10.5281/zenodo.22980503. The object of regulation—the unbounded learned function—cannot be functionallysafety-certified, because the property that would make certification possible, a totalbehavioral specification, does not exist in the object. That paper was critical. It closed a doorand named what was behind it. This paper is constructive. It presents one alternative: ST-T1024, an open standard andreference implementation for deterministic, embedded, sovereign language models. The argument proceeds in four parts. First, it recalls the fallacy and extracts from it a set of conformance criteria—the conditions any constructive alternative must satisfy to escape the category error rather than relocate it. Second, it presents ST-T1024 and shows how each criterion is met by a specific architectural mechanism: ternary arithmetic, finite-statescheduling, strict WCET enforcement, domain-tile isolation, immutable core memory, and aformally verified invariant set. Third, it addresses the central design question—what role, if any, a learned component may play in a deterministic system—and answers it with a layering principle: the learned component is permitted downstream of the decision, never inside it. The Tiny Transformer in ST-T1024 is deterministic in operation but unspecified in behavior; it is therefore confined to surface realization, where the semantic payload has already been fixed by the specified layer. Fourth, it maps the architecture to the fourfunctional safety standards that the prior paper showed exclude the stochastic paradigm, and shows that ST-T1024 satisfies them not by exception but by construction. The paper concludes that the category error is escapable, that the escape is architecturalrather than regulatory, and that the governable layer has been available all along.