A lightweight reference-free OCR quality estimator. It combines OCR-noise heuristics, lexical plausibility, and tokenizer fragmentation.