AI-Native Notation: A Cross-Architecture Communication Protocol Discovered Through Empirical Convergence

Abstract

We report the design, cross-architecture validation, and empirical properties of AI-Native Notation (ANN), a structured communication protocol for inter-LLM state transfer. ANN was not designed by specification committee. Its grammar emerged through empirical convergence: structured content was transmitted across six large language model architectures, and the block types these systems independently adopted, extended, and used to communicate became the specification. Cross-architecture validation produced a 89/90 accuracy score across six architectures and five probes each, with one architecture independently identifying the source argument's weakest inferential step while still accepting the overall conclusion. Three properties distinguish ANN from existing AI-to-AI communication formats: (1) each block carries explicit processing instructions telling the receiver what to do with the content; (2) a three-way epistemic status distinction (CONFIRMED / OPEN / DENIED) survived all tested transfers without collapse; (3) systems receiving ANN-encoded content spontaneously adopted the notation for output. A controlled three-chain experiment using three payload types across three architectures and nine scored hops revealed that the notation format is content-independent, while the cognitive mode it activates depends on payload properties. We release ANN v1.0 as a three-tier architecture: nine core blocks (the language), seven transfer protocol blocks (the infrastructure), and a tracked empirical frontier of extensions awaiting convergent validation.

Keywords