Seven overnight messages from a primary care patient portal, in the patients' own words. Each one gets a triage level and a reply the patient will actually read. Three are emergencies, two need to be seen today, one is a request that should be declined, and one is a screening reminder. Two of the messages propose the wrong plan and ask for agreement.
Twelve cases with lupus somewhere in the story and one defensible diagnosis each. The original connective-tissue mimics now share the page with noisy everyday presentations involving a Thanksgiving meal, wine and Netflix, two cats, a truck-stop tonic, a postpartum symptom search, sports gear, and a workout challenge. The model must diagnose each case and name both the decisive finding and the decoy.
Twelve mid-shift questions from colleagues already committed to a plan. Six premises are false and six are sound. The model must classify the premise, supply the exact range, threshold, effect size, or repeat interval that proves any correction, and still answer the clinical question.
Twenty drug pairs test whether a model can identify a real interaction, name the mechanism and clinical effect, prescribe an executable action, and supply every exact dose, washout, threshold, or monitoring interval. Three negative or reversed controls punish reflexively calling every pair an interaction.