instead of inferring from response length. This is less than ideal but
it fixes issue #396 and, given what appears to be MaxMind's inconsistent
response lengths, it may not be possible to do much better.