[ur-whale]

> ou can provide all of the reference implementations you want and you still will get variances.

A proper reference implementation should first and foremost come with an extensive battery of regression tests, something many a vaunted "standard" utterly fails to provide, being instead tomes upon tomes of impossible to decipher specifications, written - of all things - in human language.

Such a battery of regression tests, if properly design ought to take care of your "variations" in fairly short order.

[doug_durham]

In practice they never do. There are too many corner cases that people don't think of. It's almost a law of nature that you can't capture the complexity of any non-trivial specification in regression tests.