> If you really can't afford mistakes you have to consider the OCR inaccurate.

Isn’t this close to the error rate of human transcription for messy input, though? I seem to remember a figure in that ballpark. I think if your use case is this sensitive, then any transcription is suspicious.