Well, seems as it's my mistake, as I'm not too intimate with RMVXP's inner workings anymore... if you experienced that evasion is applied after accuracy, then that has to be true.
I can, however, clear up the 75% for you... because it actually makes much sense that it doesn't hit every fourth time. Imagine eight throws of an imaginary 100-sided dice (aka
rand(100) method call). It could look like this (each value has already been applied +1, as rand(100) returns values from 0 to 99):
Throw 1 - 23
Throw 2 - 56
Throw 3 - 49
Throw 4 - 19
Throw 5 - 81
Throw 6 - 62
Throw 7 - 74
Throw 8 - 62
Now you see, in a pretty balanced field, there's 7 out of 8 values below 75, meaning an ability will hit. Because random is random, however, it could just as well look like this though:
Throw 1 - 63
Throw 2 - 96
Throw 3 - 79
Throw 4 - 13
Throw 5 - 42
Throw 6 - 91
Throw 7 - 78
Throw 8 - 32
In this case, only 4 of 8 values are still below 75, meaning only 50% of all attempts will actually hit the target, based on a 75% accuracy. This, however, isn't very likely to happen compared to the one above, as random values tend to balance themselves out statistically (not going into detail on that, as I'm not too wise regarding the exact details either).
If you do want to go for "every fourth hit misses", that's possible to do - you just have to replace your random value with a counter variable that counts from 1 to 4 (well, 0 to 3 really). Whenever it's at any numer you choose, have it miss, whenever it's at any of the other three, have it hit. Then, increment the variable by 1 and you're ready for the next use. This is easily to script and easy to integrate, however you will find it dead boring to play with a skill system like this.
The problem with dice is that players tend to complain about random determination (as oftenly seen in Team Fortress 2, for example, where people complain about item drops being too random and unfair, for example), however it's a tedious task to mess with it. At the end of the day, dice are a good way to bring variance in a game. As an example for games that can be easily exploited because of the lack of random generators, I'd like to name Lufia 2 (which is a great game nevertheless), where everythign depends on previous actions and therefore is 100% predictable. For seeing video proof of that, check out
Dragondarch's speedrun of the game.