On the one hand, I like it. One thing I've always liked about GURPS was how the skills basically represented a shift in the bell curve.
The difference it makes in a system designed on a linear probability curve is probably too radical, though. Let's meditate on this a bit.
1d20 2d10 Armor
----------------------------
1 100% 100%
2 95% 100%
3 90% 99%
4 85% 97%
5 80% 94%
6 75% 90%
7 70% 86%
8 65% 79%
9 60% 72%
10 55% 64% Nude
11 50% 55% Padded
12 45% 45% Leather
13 40% 36% Studded Leather
14 35% 28% Chain Shirt
15 30% 21% Chainmail
16 25% 15% Banded Mail
17 20% 10% Half Plate
18 15% 6% Full Plate
19 10% 3%
20 5% 1%
Under the d20 system, Joe Schmoe with +0 BAB can hit Joe Blow with AC 10 55% of the time. Under the 2d10 system this changes to 64% -- a nine percent difference, which to put it into perspective amounts to almost a +2 BAB in terms of percent chance to hit.
Suppose Joe Blow decides to protect his assets with some padded armor. He throws a quilt over his head, say. In d20, this makes it 5% less likely that Joe Schmoe can damage him. In 2d10, it's 9%. At the very least, this suggests that the default AC of a nude palooka is going to have to be adjusted.
In d20, Joe Blow in a chain shirt is 20% harder to damage than he is in his birthday suit. In 2d10, he's 36% harder to hit. How long were you planning for combat to take?
Still, it's not a bad fit. A +1 still averages to about a 5% difference. It favors light armor, which will change the nature of the game. Difficulty checks and skills with be thrown all out of wack because skill points will be less effective at low and high levels.
I'm sure I'm forgetting to mention something.