MedPlain: Local, Bilingual Plain-Language Simplification of Clinical Text with a Faithfulness Check

Patients frequently cannot understand their own clinical documents, a health-literacy gap associated with poorer outcomes. Large language models (LLMs) can rewrite clinical text into plain language, but their fluency masks a safety-critical weakness: they can silently alter clinically material facts. We present MedPlain, an open-source system that simplifies clinical text at a selectable reading level, in English and Arabic, entirely on the user's device, and pairs each simplification with a two-layer faithfulness check (a deterministic dose/number verifier and an LLM semantic verifier). MedPlain further ingests printed prescriptions from PDF or image and applies an OCR confidence gate that refuses low-confidence reads to prevent an OCR-to-hallucination failure mode. We evaluate the system on a gold-standard set of five prescriptions spanning oncology, cardiology, and general practice, with documented "known facts" and "known traps." Readability improved from a mean Flesch-Kincaid grade of 21.5 to 7.4 and 95% of numeric facts were preserved. However, the semantic verifier detected only 1 of 24 subtle meaning-level distortions, and on a handwritten document the pipeline produced a fluent but fabricated prescription. We report these negative results transparently: numeric fidelity and fluency create false confidence, and robust semantic faithfulness for medical simplification remains unsolved. MedPlain is positioned as a research and educational prototype - not a medical device - and its design centers on privacy, honest safety framing, and mandatory human review. Code and data are openly available.