From Plausible Narrative to Sound Abduction: A Governed Abductive Architecture for Medical Digital Twins in Multiple Sclerosis Care

Large language models (LLMs) embedded in medical Patient Digital Twins (PDTs) exhibit a systemic vulnerability: they can generate fluent, narratively persuasive, yet abductively unsound clinical explanations. In this context, abductive soundness means that an explanation preserves mechanistic plausibility, temporal coherence, explicit handling of missing premises, and sensitivity to counter-evidence. This article reframes the problem as architectural rather than as a mere deficit in training data. We identify three recurrent modes of abductive failure — missing-premise neglect, weak-mechanism support, and counter-evidence discounting. They arise when local semantics, formal world ontology, and the role-specific clinical semiosphere are collapsed into a single surface flow of generation. We propose a governed abductive architecture organised around seven runtime contours and operationalise it in the MS-AGIP platform for multiple sclerosis care. The architecture separates three subsystems: an ontology-guided Research Framework, a clinician-facing Neurologist Digital Twin, and a patient-controlled Patient Digital Twin. We show how disease-specific causal templates, evidence-tiered biomarker reasoning, provenance labels, temporal-coherence checks, molecular-clinical discordance detection, and governed patient-feedback updates jointly transform plausible narrative into sound abduction. The article presents an architectural blueprint and validation protocol aligned with TRIPOD+AI and DECIDE-AI. The architectural-versus-scale distinction has direct implications for safe medical AI: the difference between a fluent and a sound clinical system lies more in architecture and governance than in model size. None of the subsystems has yet been clinically deployed.