Evaluating hard-constraint reliability in clinical nutrition planning: a feasibility-aware benchmark of large language models and knowledge-based systems

Preprint, version v1-method-preprint. Not peer reviewed. This record contains the main manuscript and the accompanying Supplementary Methods and Reproducibility Appendix (S1–S22). Clinical nutrition plans may appear nutritionally adequate overall while still violating a mandatory clinical or nutritional constraint; conversely, failure to generate a plan does not establish that no admissible plan exists. This study develops a feasibility-aware hard-constraint benchmark for constrained clinical nutrition planning and evaluates its exploratory behavior across Metabolic Nutrition Pro (MNP) and three productized large-language-model API configurations. Benchmark A evaluates end-to-end planning, whereas Benchmark B provides the same eligible recipe pool to all systems. The exploratory calibration used 10 protocol-generated clinical scenarios. After correction of identified benchmark defects, scheduled-run valid-plan rate (VPR) was 0.90–1.00 on Benchmark A and 0.50–0.70 on Benchmark B. Structured input, closed rules, runtime validation, and one informed repair did not eliminate hard-constraint failures in the evaluated LLM configurations. An archived exploratory MNP run also returned success despite violating an encoded potassium hard constraint after an unsuccessful admissible solve. The study is methodological and exploratory. It does not establish comparative superiority, clinical safety, or clinical effectiveness. The prospectively specified confirmatory campaign has not been executed; independent blinded clinical validation and external source adjudication remain pending. Proprietary MNP implementation details, precomputed commercial parameter files, and other commercially sensitive materials are not included. Releasable machine-readable study artifacts are retained for request-based audit as described in the manuscript.