Hoo, boy.
*Pops her knuckles.*
Here we go.
Statistics is not math.
Statistics is a logical method that incorporates math as a principal tool. You can do the math flawlessly, but if you select a nonsensical model or fail to interpret data sensibly, the math becomes useless.
I find your math sound and your statistical analysis to be an abject failure in model design and interpretation. Good math plus bad logic leads to bad conclusions.
You have correctly calculated the probability that increasing AP values from one to four will prevent at least one success, and assume in your model that preventing two, three, or four successes is a distinction not worth accounting for.
Let us call preventing a success our event. If your AP manages to prevent a single success, then 100% of an event has occurred. If you prevent two successes, 200% of an event has occurred! You can have percentages greater than 100%, depending on your model. You just have to interpret what that percentage means. That 133 1/3% does not mean that if you roll three times, there will be four separate rolls in which an event occurs. It means on average, 400% of an event shall occur overall because an event can occur more than once per roll.
By your own math, of AP -4 having an 80.25% chance of negating at least one success, a 40.74% chance of negating at least two successes, an 11.11% chance of negating at least three successes, and a 1.23% chance of negating four full successes, you agree with me. Once you count every event, that 80.25% where a first event occurs, plus the 40.74% where a second event occurs, plus the 11.11% where a third event occurs, plus the 1.23% where a fourth event occurs? That's an average of 133 1/3% of an event per roll!
Here is a big spreadsheet full of lots of random numbers and some math.What I did here was assume a Body 4 human wearing an armored jacket. I used randbetween(1,6) to model a 6-sided die, spread across 16 rolls, then columns all the way out to column ZZ.
Below that, I constructed a sum of if(CELL>4,1,0) statements adding up all 16 random die rolls in the respective columns.
Below that, I did the same, counting the top 15 dice, then 14, then 13, then 12 dice.
On the far left, I took the mean of each set of randomly rolled successes for each number of dice rolled.
The results?
16 dice: 5.35 successes on average.
15 dice (AP -1): 5.021429 successes on average.
14 dice (AP -2): 4.698571 successes on average.
13 dice (AP -3): 4.361429 successes on average.
12 dice (AP -4): 4.06 successes on average.
In other words, what I've been saying all along. Each additional point of AP reduced average number of successes by approximately one third of a success. Each point of Penetrating Strike was equally valuable, as the goal of the ability is to reduce the average number of successes your opponent scores on soak rolls. No diminishing returns.