NeuroGlyph: A Semantic Interlingua Protocol over Natural Semantic Metalanguage Primes — Specification, Reference Codec, and Fidelity Evaluation

Abstract

We present NeuroGlyph, a semantic interlingua protocol that encodes the meaning of human and agent communication as a rooted directed acyclic graph over the 65 semantic primes of Natural Semantic Metalanguage (NSM, Chart v15, Goddard and Wierzbicka). The protocol introduces an explicit pragmatics layer (intent, mode, envelope), a 10-axis diacritic feature mask, and a lossless macro-crystallisation mechanism, all serialised deterministically so that the same meaning always yields the same canonical hash. We release a reference codec (rule-based encoder, graph decoder, structural verifier) and evaluate it on a 48-phrase corpus plus a 10-phrase held-out set. The codec covers 41/48 phrases with a mean weighted reconstruction loss of 0.002, passes 4/4 adversarial paraphrase tests (including 2 cross-lingual pairs) and 3/3 negative semantic contrast tests, and yields identical canonical hashes for a Russian and an English decoding of the same gold graph when both re-encode cleanly. We also describe our anti-Anglocentrism methodology: a 25-language exponent corpus (14 languages at full coverage), an expressibility audit that detects ontology-level gaps against the exponent corpus, and a typological canary set for non-Indo-European languages. Finally, we specify and implement NG-BIN v1, a binary wire format for the glyph graph, and the ng.message binding that carries it over JSON-RPC and over the Model Context Protocol; on the 58-graph gold set the binary form is 9.46x smaller than the equivalent schema-message JSON. NeuroGlyph is not a replacement for natural languages nor a machine-translation system; it is a semantic substrate that any message can carry alongside its surface form.