Just to add to that: Coggan alone has changed his mind on “the” definition of FTP over the years. Machine Learning (the way it is used now) wasn’t a thing nor was there a corpus of data comparable to what TR has, and a “definition” via ML algorithms wasn’t on anyone’s radar, because it wasn’t possible. That doesn’t mean a statistical approach based on a data set that an exercise scientist couldn’t even dream of creating can’t be more accurate or better.
Even within the “accepted definitions” — or rather, most commonly used FTP test protocols, there is a lot of variability.
You are assuming that AI FTP has a tendency to overestimate “FTP”. How do you know that? According to @Nate_Pearson about 2/3 will see a decrease in predicted FTP and less than 20 % an increase or a significant increase. There are several threads and posts like this where users see a decrease of AI FTP.
Importantly, a smart choice by TR was to simply make TR AI oblivious of all of this, because according to TR they are based on absolute power numbers and are not anchored in FTP.
I don’t think it is wise to use several training methodologies simultaneously. TR hasn’t used ATL and CTL ever (as far as I can remember at least), whereas other training approaches use CTL and CTL ramp rates, something that is quite a bit more primitive than what TR is using at the moment.