Deterministic Context Transaction Protocol (DCTP): A Governance Layer for Confidence- Gated, Versioned Context Exchange in Multi- Agent LLM Systems

Multi-agent large language model (LLM) systems increasingly coordinate several cooperating agents to complete a shared task, but conventional designs let each agent reinterpret upstream information freely, producing inconsistent field values, propagated hallucinations, and workflows that cannot be reproduced after the fact. We present the Deterministic Context Transaction Protocol (DCTP), a middleware layer in which agents exchange versioned, confidence-scored Context Transaction Objects (CTOs) rather than free-form text. Each CTO passes through a governed lifecycle - including an immutable LOCKED state reached only when a confidence threshold and dependency conditions are satisfied - and can be modified only through a validated update transaction rather than direct write access. A dedicated arbitration component resolves disagreements between agents using source-reliability and historical-accuracy weighting, and a replay component reconstructs prior executions from the preserved transaction log. We situate DCTP against four related bodies of work: classical optimistic concurrency control and workflow pre-locking in database systems; industry agent- communication standards (Anthropic's Model Context Protocol and Google's Agent2Agent protocol); a patented confidence-gated agent context space (WO2021084510A1); and concurrent 2025–2026 academic proposals for confidence-calibrated, trust-weighted multi-agent context management. We find that DCTP's individual mechanisms each have close antecedents, and that its main contribution is architectural: combining dependency-graph-gated locking, propose-only agent interfaces, and model-version-linked replay into a single general-purpose protocol rather than a domain-specific instance. We report this novelty assessment candidly, including where the contribution is narrower than initially believed.