Bringing some clarity into the debate over the directionality of grammaticalization presupposes that the concepts involved be first explained. First of all, the methodological standards for empirical proof in a science like linguistics have to be made explicit. The question is: How do we prove the diachronic sequence of the phases that a process has run through in a particular case; and on which basis can we then generalize on the direction that this type of change generally takes? Applying these questions specifically to grammaticalization, we are asking: What is the empirical evidence for the course of grammaticalization?

For a set of data to count as evidence for grammaticalization, it must fulfill three conditions:

1. There are two historical stages of language L, earlier L1 and later L2.
2. L1 has form F1 and L2 has form F2, such that F2 is diachronically identical with F1.
3. F2 is more grammatical than F1.

From this we may conclude, e contrario, that certain sets of data that have been adduced as illustrating grammaticalization do not count as empirical evidence:

• Synchronic data – like the grammaticalization of the copula in Jula, to name but one example – do not count as evidence. Their interpretation as cases of grammaticalization may be flawless, so they may be perfect illustrations of grammaticalization; they just do not count as proof.
• If either or both of the two stages L1 and L2 is reconstructed (cf. the Mayan passive suffix), the data do not count as proof. Since it is grammaticalization theory which renders the reconstruction plausible, adducing a reconstruct as empirical evidence of grammaticalization is circular.
• If F1 and F2 are not diachronically identical, none can be grammaticalized from the other. The Latin future cantabo and the Spanish future cantaré are isofunctional and diachronically related, but not identical. They may be assessed as grammaticalized to an equal degree, but they do not instantiate one grammaticalization process.

The net balance of this is: A case that falls short of any of the three requirements may be a perfect example of grammaticalization; it just does not count as empirical proof.