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

Multi-agent AI systems lack a coordination format. Current methods rely on prompt engineering that does not generalize across architectures. This paper introduces AI-Native Notation (ANN), a structured format that encodes processing instructions at the block level. Each block carries a per-message contract specifying how the receiver should handle each content component. A cross-architecture replication study tested ANN across eight large language models, covering both Mixture-of-Experts and dense architectures, with four tested via consumer interfaces and four via the Together.ai API. Structural adoption was uniform: all eight models honored the block-level format on first contact, without prompt engineering. A format comparison experiment isolated the format effect from content: only ANN produced uniform adoption across the four models tested, while equivalent content in JSON, XML, and structured English did not. A nonsense control isolated the format effect from semantic content: structural adoption persisted with fabricated content. A three-chain experiment isolated format propagation from content propagation across multiple receiving systems. Two pre-registered experiments extend v7.0.2: a tag-shuffle ablation falsifies the specific-tag-token hypothesis and locates ANN's contribution at the schema level as block-tagged structured notation with processing instructions in surrounding prose; a theorem-proving handoff pilot reports a pre-registered null at N = 100, with conclusion correctness saturating across all five format conditions, indicating a task-difficulty ceiling rather than evidence against the coordination hypothesis. The empirical record positions ANN as a candidate coordination layer for AI-native software ecosystems (Russo, 2026) and as a candidate inter-agent communication format for the Software Engineering 3.0 vision (Hassan et al., 2026). The format specification, validator, replication scripts, and experimental record are released under permissive licenses (CC BY 4.0 for the specification, grammar, probes, and paper; Apache License 2.0 for validator and analysis scripts).