Applying a bell curve to the base d20 mechanic is a rather foolish idea. Here's why.
1) As pointed out, you remove the importance of crits, making all weapons essentially the same and neutering spells based on crits, like keen weapon.
2) The importance of spells that grant bonuses to die rolls or DC's (heck, that's pretty much all the spells that don't do damage) is vastly magnified because the d20 system. In the d20 system a +1 bonus is just that regardless of what the DC previously was... Each +1 or -1 is a 5% change in probability. On a bell curve producing a change of the same amount can either have little to no effect (near the middle of the curve) or can have such a massive effect as to throw the spell out of whack.
3) For the above reason, allowing ANY MAGIC to stack is broken with a bell curve. And all higher level spells which can impart 10 point or more swings to a roll are likewise quite broken.
If you want bell curve like effect SET YOUR DC'S APPROPRIATELY...
You can also use open ended rolling, which works as follows. If a natural 20 is rolled, instead of it being an instant hit / success just allow a reroll with a +20 bonus. If a natural twenty is rolled twice, then roll with a +40 bonus and so on. This makes a DC 50 test about as difficult as rolling a 20 using some of the methods offered above, but it is *possible,* and for epic characters it is even *likely* (<sarcasm> Gasp.... A heroic character doing something heroic.. Who'd thought? </sarcasm>)
Fumbles can be done the opposite way.. Instead of a natuaral 1 missing, just roll again with a -20 penalty for each reroll.
A second way is to count sucesses against a test, referred to as a cascade check in my upcoming book Role & Rule. Here you roll a d20 check as normal against a DC. Each time you hit it, you get to try again. Once you miss, you stop rolling and check to see how many times you hit the DC. Most frequently used with checks where both success and quality of success is important to measure, cascade checks are used most frequently with knowledge checks. 1 success gives general knowledge about the question, but no details. 2 success gives more specific knowledge, with major points missing, inaccurate or obscured. 3 successes give detailed information with minor points innacurate and 4 successes give complete information. In this system a getting multiple checks drops off rapidly the further the DC moves away from the number of ranks the character has invested in the check.
For instance, an open question posed to a character with 10 ranks in the appropriate knowledge skill. Complete and accurate information on a DC 15 check is much easier than even a DC 20 check, and certainly easier than a DC 30 check, which lies on the fringe on their knowledge base. I'd like to see a graph of this plotted, by one of those who like to do such things...
. 3d6? Even Nah-er, can't roll nat1s or nat20s. So how bout this: Median Roll. IE, roll 3d20, take the median.
