You're asking if you can predict how average human beings would transmute a given word based on a set of transmutation rules, as far as I can understand.
For such a prediction to occur, you need either an evidence-based prior assumption or a predictive model that has been evaluated and whose scope comprises your specific question. The first road is the one of previous breaches, from which you can infer for the same alphabet and rules how words will be transformed. You can then use this information to describe the actual entropy of a password input space -- rather than a "set" entropy that does not take into account the probability distributions of words. In practice you won't have enough data to cover the whole input space of an alphabet so such a metric still relies on some level of extrapolation (i.e. "all the unobserved patterns are equally likely to occur"). Bonneau has a paper on security metrics for passwords that better reflect the reality of how password spaces are used by end users (possibly this one).
The second road would require you to do quite a lot of experimenting with human subjects and devise general rules of how people choose to transmute words based on available symbols. You would do multiple experiments to assess which symbols are preferred, how they are placed, where, and you would need to cover an ecologically valid set of:
- input alphabets
- proposed transmutation symbols and rules
- interfaces to password authorship and entry
- social settings (this one just defeats the astonishing majority of security researchers doing human subject research)
It'd be extremely hard and costly to make such a model complete (i.e. accounting for all possible types of transmutations), valid and reliable.
Especially, it is my intimate belief that interfaces have more impact over how people authenticate than passwords (but I have neither the time or motivation to argue this in any scientific way :-) ), and I would be willing to bet that if such a model were made for e.g. desktop computers with a certain type of keyboard it would fail for e.g. tablets with a different keyboard layout (for instance if "@" is available with a lesser interaction cost than "1" on a keyboard it may be preferred for that very reason).
Thinking of this problem from a UX perspective simplifies the matter a great deal: why do people always pick very basic transformations e.g., add a number and a symbol at the end of an existing password? Because it was cheap and easy to do so on an existing password. What did we learn from this? Either we make all options equally painful to the user so that they do randomly pick a password, or we stop asking them such a ridiculously difficult task of selecting random but memorable auth factors.