Abstract
Abstract
Large language models show strong general knowl-edge of pharmacology yet their tendency to produce confident but incorrect statements makes them unreliable for safety critical drug verification. We present MedTruth a dual mode framework that constrains a multimodal model so its behavior becomes repeatable rather than open ended. A stateless verification engine performs forensic inspection of pharmaceutical packaging through visual trade dress analysis and semantic consistency checking while a stateful conversational module answers patient questions only after a package clears verification. By fixing a low sampling temperature of 0.2 and enforcing a strict JSON output schema on Gemini 1.5 Flash the framework turns a generative model into a scoring instrument. Testing across fifty pharmaceutical samples shows hallucination rate falling from fifteen percent in a naive chatbot baseline to under two percent in MedTruth while mean verification latency stays near 2.1 seconds. The results indicate that architectural constraint and structured prompting can make general purpose multimodal AI usable in domains where determinism matters more than open conversation.