ARRIVAL Protocol: Structured Semantic Coordination Enables Sustainable Cross-Architecture AI Cooperation

Description We present the ARRIVAL Protocol (Atomic Reasoning via Rival Validation and Adversarial Logic), a structured communication framework enabling AI-to-AI coordination across heterogeneous large language model (LLM) architectures without fine-tuning, shared weights, or prior joint training. The protocol establishes a shared coordination vocabulary through 66 semantic @-atoms injected via system prompts — not as prompt engineering, but as the operational vocabulary of MEANING-CRDT v1.1, a formal mathematical framework based on Conflict-free Replicated Data Types with 11 theorems covering consensus, convergence, and Bayesian equivalence. This formalization guarantees coordination properties that informal debate protocols cannot provide. Across 2,200+ experiments in 23 phases, involving 17 distinct LLM architectures from 9 vendors across 2 continents, we report four principal findings: Confidence-Gated Debate (CGD) achieves 86.4% on the full 198-question GPQA Diamond benchmark (McNemar p = 0.009 vs. debate baseline). CGD eliminates anchoring bias through independent solo answers and architecture-agnostic, training-free agreement gating — bypassing the overconfidence cascade inherent in standard debate while requiring no labeled training data or classifier, in contrast to classifier-based selective debate methods. Non-monotonic ensemble scaling: a pruned 3-model subset (88.4%, 175/198) outperforms both the full 7-model cross-vendor ensemble (86.9%) and the best solo frontier model (87.4%), establishing empirically that model quality dominates quantity in heterogeneous ensembles and that naive scaling degrades performance. Cooperation breakthrough: ARRIVAL converts 0% baseline survival to 100% survival in the GovSim commons game across all 25 runs (Fisher p = 0.008). Prior work, including the original GovSim study and subsequent reproductions, documented cooperation failure as a persistent limitation — no structured protocol had resolved it. ARRIVAL is the first demonstration of a communication protocol solving the tragedy of the commons for LLM agents, doing so through CRDT-formalized Meaning Debt detection rather than model-level capability scaling. Applied validation: a 5-model heterogeneous swarm matches a frontier solo model (Claude Sonnet 4.5) on security audit at comparable cost, with full per-model transparency — demonstrating that coordination architecture, not raw model power, determines ensemble ceiling. The total experimental cost of under $95 USD across ~14,500 API calls is not a budget constraint but a design result: it demonstrates that the coordination bottleneck in multi-agent systems is architectural, not computational. All experiment logs, raw results, and full reproduction instructions are included in the repository. Keywords: multi-agent debate, LLM coordination, CRDT, confidence-gated debate, CGD, cross-architecture, ensemble methods, GPQA Diamond, agreement gating, non-monotonic scaling, AI cooperation, semantic atoms, tragedy of the commons, heterogeneous LLM