Reads garbled messages (speech-to-text, autocorrect, swipe, OCR) the way they were meant, shows the reading before acting, and asks when ambiguous.