The artificial intelligence industry describes its learnable object by a single formula: a conditional distribution optimized with respect to output statistics. We fix this class verbatim — language, diffusion, and RL models — and call it the stochastic paradigm. That same industry has operationalized its goal by a canon of ten cognitive domains; the deficient domains of that canon decompose into four operations: inferring and applying a rule, reusing a premise, intervention and counterfactual, consolidation of what has been inferred into an object. The work poses a single question — are these operations expressible in the internal language of the paradigm itself, the language in which its systems build the complex out of the simple — and answers it with a theorem. The answer is negative by equivalence, not by deficiency; the engine of the proof fits into one line: the naturality of copying, read for a Markov kernel, yields p = p² for the probability of every event, that is p \0, 1\ — only the certain is copyable. The structures behind these operations — copying, projection, rule-as-object — are the building material of cartesian categories and are therefore not «not yet learned»: their appearance in the probabilistic layer is equivalent to the destruction of the defining property of the paradigm itself — stochasticity. This boundary we call the cartesian ceiling: it belongs not to an architecture and not to a generation of systems, but to the paradigm itself, and scaling moves along it without raising it. Achievements in reproducing values are untouched by the theorem; a second, independent line shows that a deterministic unrolling likewise reproduces the values of the operations without possessing the structure. Precisely this difference is measured by isolating benchmarks: a model explains a concept in 97. 7\% of cases and applies it in 67. 9\%, while every new test in the ARC series finds the frontier at zero until the next layer of external engineering. Both lines explain the documented jagged profile: strength where values suffice, failures where the operation must exist as an object. Impossibility results from three unrelated disciplines have landed on the same boundary — Pearl's causal hierarchy, the TC⁰ limit of transformers, the statistical lower bound on confabulations — and our construction allows their enumeration to be read as a map of one boundary. Engineering scaffolds attempt to realize the operations in a deterministic layer and thereby confirm the theorem rather than refute it. The theorem is unconditional; the cognitive reading is conditional on an explicitly flagged thesis about formalization. The diagnosis is constructive: the theorem itself names the layer in which the required structures live, turning a prohibition into a technical specification for an ascent up the categorical ladder.
Egor Vikhlyaev (Fri,) studied this question.