One patient has more than one problem. Their food advice should know that.

There is a strong diabetes guideline, a strong kidney guideline, and a strong blood-pressure guideline. Each one is written as if it is the only thing going on. A real patient often has three or four at once, plus the medicines that treat them.

When you stack that advice, it can collide. A diabetes plan pushes beans and greens for fiber and potassium. Reduced kidney function plus an ACE inhibitor pulls potassium the other way. Reading all of it correctly, for this patient, in a short visit, is hard. Most tools do not even try.

A large language model writes fluent diet advice, but it can invent a number and miss a drug or kidney interaction. For food as medicine that is not a tone problem, it is a safety problem. A wrong potassium target for a patient on an ACE inhibitor is the kind of mistake that matters.

So we drew a hard line: the model never writes a medical fact, a number, or a food instruction.

The peer-reviewed record now says the same. Testing language models on type 2 diabetes meal planning, a 2026 study concluded that “inconsistencies in energy and nutrient content, together with the lack of consideration of individualized requirements and comorbidities, demonstrate that AI-based dietary plans do not yet provide a level of accuracy and appropriateness sufficient to replace dietitian-guided nutrition therapy” (Healthcare 2026;14:739). A 2025 review of language models in clinical nutrition warns that hallucination means “the model generates plausible-sounding but incorrect or nonsensical information” (Front Nutr 2025;12:1635682). And when models were tested against cardiovascular dietary guidelines, “the off-the-shelf models scored lower on all measures and produced some harmful responses” (J Med Internet Res 2025;27:e78625).

Every recommendation lives in a curated evidence bank, and every entry carries its source with a verbatim quote and a link. A plain, deterministic engine reads the patient and the bank and assembles the plan. If a recommendation has no source, the build fails and it never ships.

The result is checkable and repeatable. The same patient gets the same plan every time, and you can click any line back to the guideline or the USDA food record behind it.

The engine pulls every rule that applies, then merges the targets with the safest one winning: the lowest ceiling, the highest floor. It runs the drug and disease safety checks, swaps foods rather than dropping a food group when it can, and shows every conflict with which rule won and why. Nothing is hidden and nothing is improvised.

The tool does not diagnose, change medicines, or handle acute care, and it stays inside the conditions it has been built for. When a case calls for a clinician, such as pregnancy, advanced kidney disease, or a high-risk medicine, it refuses and routes out rather than guessing.

It is decision support with a clinician in the loop. The evidence bank is clinician reviewed, newly added conditions are marked pending a final sign-off, and HIPAA and device-classification questions go to compliance before any real pilot.

Few institutions can ground food-as-medicine advice in the Adventist Health Study and a century of plant-forward practice. This puts that authority into a working tool a clinician can hand a patient, and measures whether it helps.

Lifestyle medicine treats chronic disease with daily habits. The American College of Lifestyle Medicine frames it as six pillars: nutrition, physical activity, sleep, stress, connection, and avoiding risky substances.

This tool goes deep on one of them, nutrition, and does it safely for a patient who has several conditions at once. It is built to sit alongside the rest, not replace them: the clinician still owns activity, sleep, stress, connection, and substance use. The dietary approach rests on the Adventist Health Study, Loma Linda's decades-long cohort linking plant-forward eating to better outcomes.

Open the pillars scorecard to rate a patient across all six.