Thanks Stolfi
My point was not about raw character frequencies! It was about a conditional dependency across a visible token boundary: the internal form before final y changes the probability that the next token begins with qo. And we are talking about more than 1,000 cases...
Sorry, but that cannot be dismissed as ordinary character statistics.
If qo is “and”, the distribution should mainly follow syntax or enumeration structure, not the internal shape of the previous token!!!
If y is a pitch marker, the model still has to explain why the material before y changes the following qo probability so strongly.
Sorry, but the question remains: what is the concrete mechanism that connects the internal pre-y form with the next token’s qo?
My point was not about raw character frequencies! It was about a conditional dependency across a visible token boundary: the internal form before final y changes the probability that the next token begins with qo. And we are talking about more than 1,000 cases...
Sorry, but that cannot be dismissed as ordinary character statistics.
If qo is “and”, the distribution should mainly follow syntax or enumeration structure, not the internal shape of the previous token!!!
If y is a pitch marker, the model still has to explain why the material before y changes the following qo probability so strongly.
Sorry, but the question remains: what is the concrete mechanism that connects the internal pre-y form with the next token’s qo?
![[+]](https://www.voynich.ninja/images/collapse_collapsed.png)