Beyond Hallucination: A Neuro-Symbolic Framework for Fidelity-Preserving Clinical Interpretation in Low-Infrastructure Settings

Abstract Background: Large Language Models (LLMs) hold promise for clinical applications, but existing evaluation methods focus primarily on accuracy and hallucination detection. These performance-centric metrics fail to capture silent failures, such as information omission, inconsistency, or the inability to adapt to evolving clinical guidelines, especially in patient-facing medical summaries. Objective: We introduce NXAiR Pipeline (Neuro-Symbolic Executable Arbitration for AI Reasoning), a lightweight neuro-symbolic framework that operationalizes structural fidelity, ensuring all authorized clinical information is accurately rendered in patient-facing narratives. To evaluate system behaviour beyond traditional performance, we propose TREAD-A, a governance-oriented framework measuring Transparency, Reproducibility, Explainability, Auditability, Determinism, and Adaptability. Methods: Clinician-generated structured token maps (STM, representing the authoritative clinical state) served as the deterministic source of truth. NXAiR Pipeline translated these into natural-language summaries without probabilistic reasoning. Pilot evaluation included 27 clinician-reviewed cases and 32 patient comprehension assessments. Outcomes were measured against TREAD-A attributes. Results: NXAiR Pipeline achieved 0% omission and addition rates, with perfect reproducibility. Transparency and auditability were ensured via JSON-linked STM tokens. Real-time updates to clinical instructions demonstrated adaptability without retraining. Patient comprehension of diagnoses improved from 18.8% to 87.5%, therapeutic understanding from 12.5% to 93.8%, and dosing adherence from 37.5% to 87.5%. Conclusion: NXAiR Pipeline operationalizes structural fidelity in clinical AI, moving evaluation beyond traditional accuracy or hallucination metrics to architectural trust and operational safety. TREAD-A provides a concrete framework for governance-oriented assessment of patient-facing medical summaries. This paradigm offers a scalable solution for resource-constrained environments and sets a new standard for safety-critical AI systems. Provisional Disclosure: The NXAiR™ framework was previously introduced in a conceptual White Paper (doi: 10.5281/ZENODO.18207434)) and is the subject of a pending provisional patent application. This manuscript presents the formal scientific evaluation, the TREAD-A governance framework, and comparative performance outcomes. It is intended as a high-level architectural discussion; no proprietary implementation code is disclosed herein.