The algorithm can be modified to handle bad "length" values by adding some logic in the place where it currently dies. If you find a bad record, you could rewind to the last known "good" position via tell and seek and look for the next "==" (an alternative might be to implement your own buffer to look around in, but read should already be buffered). Similar logic would allow you to find "==" that are not aligned properly - the example above doesn't handle the case of the file starting with "x==". (Yet another approach is to read the file byte-by-byte instead of every two bytes - then you would extend your states and have "expect_first_equals", "expect_second_equals", "expect_first_length_byte", and "expect_second_length_byte".)