Data on a document is recognized using at least two machine recognition processes. Data from one machine recognition process is used as reference data; data formed using the other recognition process is used as verification data. If the verification data matches the reference data, machine recognition may be verified. If the verification data does not exactly match the reference data, an assessment is made of the likelihood that the verification data is sufficiently close to the reference data to suggest an error in the verification data. This may be done by applying a fitness function to the verification data, to assess the likelihood that the verification data represents a mis-recognized version of the reference data. In one embodiment, the verification data is OCR data, and the reference data is MICR data.