Evolution's Morse Code: Language, the cost of the uncollapsed, and the one distinction it never had to encode

Abstract

Biology evolved to identify, not to deliberate: identification is cheap and terminates, while processing is expensive and does not. Language broke that economy by granting access to states that have not resolved — the future, the absent, the counterfactual, the self as a third-person object — and the capacity was paid for in a currency the organism has been spending ever since. This essay treats natural language as an evolutionary scaffold rather than a defect: a low-bandwidth, noise-tolerant, minimal-infrastructure protocol, the only channel available, which co-evolved with the processor it loaded. It then examines what that protocol never had to encode. Every grammar forces its speakers to declare some things and lets them leave others to context. Tense is obligatory in most languages; evidentiality in roughly a quarter of them. No natural language obligatorily marks the resolution at which a claim is evaluated — because for most of human history the grain of evaluation was fixed by the body and never varied. Language models trained on that protocol appear to inherit both its map and its blank spaces. In a preregistered benchmark of four hundred items built on dated real measurements, and run against several frontier models from more than one provider, failures do not distribute like a competence deficit. Where the protocol obliges a declaration, models handle the distinction and differ from one another. Where it does not, they fail together, at the ceiling, with no spread between architectures. The essay distinguishes two kinds of blank — a permission problem, which an output schema recovers, and an absence, which it does not — and argues that the second is what a formal substrate would have to address. A falsifiable prediction is stated, together with the rival explanation the author cannot currently rule out.