So it wasn't just me lol...
Seriously, though, appreciate your attempt to put what they wrote in English lol. A few comments:
I lumped all these first comments together...
"But that doesn't actually REALLY take into context the game. So, if they highest probability assigned by the model is a slant (again, for example), but in the actual game, the DB shaded inside and covered it like a blanket, and the QB went to another read, the QB still gets dinged in that situation because he didn't hit the highest probability completion....Or, it's possible the slant, which might be the highest probable completion on a given play, the CB comes up and jams the receiver and takes the inside away on a given play, the route is dead. But there's no way the model knows this....I just thought of another flaw: this also doesn't take into account (because it can't) any pre-snap reads which are also QB decisions. "
- Did they not use any of their personal observations in the equation? I thought (or assumed) they did...because if so, they would have seen things like the above before entering the data into their program (guess they could also do it afterwards but that would be a **** and a half to do)
"Another example is the slant could actually be used as a decoy in the given play, and the actual play could be to try and leak the TE behind the slant, but there is no way for the model to know that."
- Or for them to know that lol (one of my criticisms of PFF's normal grading system)
"But the whole concept of the study is flawed right from the jump when you try and put a probability on a route depending on down/distance and game situation. You just CAN'T do that. I also don't know if they took into account things like weather, home/away, indoor/outdoor, and offensive skill position strength, what routes the QB throws best, what the receivers run best, etc."
- I seem to recall them mentioning they took into consideration offensive skill position strength (or at least the skill level of individual offensive players)