This record contains the preprint of the research paper: "From Parsing to Adjudication: A Reproducible Protocol-by-Artifacts Framework for AI Agents." Modern AI agent ecosystems rely on JSON-based protocols for tool invocation and artifact exchange. In many practical systems, protocol interoperability is implicitly validated through permissive parsing or documentation interpretation, which can lead to inconsistent behavior across implementations. This work introduces the Agent Object Protocol (AOP), a protocol-by-artifacts framework that transforms protocol specifications into executable adjudication artifacts. Instead of relying solely on schema validation or parser acceptance, AOP combines schemas, conformance fixtures, semantic checks, and CI validation gates to enable deterministic PASS/FAIL protocol adjudication. We evaluate AOP across five research questions covering parse-only validation sufficiency, deterministic replay, adversarial artifact rejection, synthetic-scale classification fidelity, and cross-implementation consistency. Experiments demonstrate that parse-only validation rejects none of the evaluated artifacts, while AOP deterministically rejects adversarial inputs and maintains consistent outcomes across independent validators implemented in Node.js and Python. The artifact package accompanying this work includes executable evaluation scripts, conformance fixtures, and CI pipelines that enable reproducible protocol adjudication experiments. This approach shifts protocol validation from documentation-based interpretation toward machine-verifiable adjudication evidence, supporting more reliable interoperability in AI agent ecosystems. Keywords: AI agents, protocol verification, artifact evaluation, reproducibility, deterministic validation, protocol adjudication.
Zhang Bin (Thu,) studied this question.