My choice of stability test is Intel Linpack, although it probably doesn't work so well on AMD CPUs; and my rule is that a system which can't survive at least 24h of Linpack is not stable --- I've seen systems fail it with stock voltages and frequencies.

I used to overclock, but stopped after realising the errors that occurred with Linpack; it really made me appreciate the fact that a CPU could miscalculate, and a system that appears "stable" can actually be silently corrupting data.