There is empirical evidence against the MAD hypothesis.
During the Cuban Missile Crisis, a Russian submarine believed that nuclear war had broken out. Three officers on board the submarine were authorised to unanimously launch a nuclear torpedo. An argument broke out among the three, in which http://en.wikipedia.org/wiki/Vasili_Arkhipov was against the launch, preventing a nuclear missile from being launched (and presumably a nuclear retaliation would not be unlikely).
Let's assign a probability that one of the officers would be in favour of launching a missile. Given that 2/3 in our sample were in favour of launching the missile, let's assign the probability to be 2/3 for any single officer. Therefore, the chance that a missile would have been launched is (2/3)^3 - 27%. Even if the probability was less than 2/3 (and people's opinions are interdependent - one person could convince another, or one person could take a contrarian stand), nevertheless it still could have happened. And this wasn't the only time that there almost was a nuclear war. (Incidentally I believe these cases show that an individual effort can at certain points have a huge impact - proof of both the butterfly effect or the 'great man theory' - as long as you redefine 'Great Man' to be a person in the right time and place).
As to the question of whether MAD worked - well it can be certainly argued to have helped, but it demonstrably did not prevent circumstances that could lead to a nuclear war.
" but it demonstrably did not prevent circumstances that could lead to a nuclear war." Do you mean "it didn't eliminate all circumstances", or do you mean "there were no circumstances that it prevented"?
The standard view of Mutually Assured Distruction (MAD) is something like:
Occasionally people will reply with an argument like:
This is an anthropic argument, an attempt to handle the bias that comes from a link between outcomes and the number of people who can observe them. Imagine we were trying to figure out whether flipping "heads" was more likely than flipping "tails", but there was a coin demon that killed everyone if "tails" came up. Either we would see "heads" flipped, or we would see nothing at all. We're not able to sample from the "tails: everyone-dies" worlds. Even if the demon responds to tails by killing everyone only 40% of the time, we're still going to over-sample the happy-heads outcome.
Applying the anthropic principle here, however, requires that a failure of MAD really would have killed everyone. While it would have killed billions, and made major parts of the world uninhabitable, still many people would have survived. [1] How much would we have rebuilt? What would be the population now? If the cold war had gone hot and the US and USSR had fallen into wiping each other out, what would 2013 be like? Roughly, we're oversampling the no-nukes outcome by the ratio of our current population to the population there would have been in a yes-nukes outcome, and the less lopsided that ratio is the more evidence that MAD did work after all.
[1] For this wikipedia cites: The global health effects of nuclear war (1982), Long-term worldwide effects of multiple nuclear-weapons detonations (1975). Some looking online also turns up an Accelerating Future blog post. I haven't read them thoroughly, and I don't know much about the research here.
I also posted this on my blog