Comment by 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.