I set up a test case vs. the AI, with Napoleon vs. Mack, and replayed the battle a number of times. Each battle should count as 2 trials, since each side has a leader. First I tried it 20 times, and had no casualties. Then I tried a further 20, and had 1 casualty. So that's 1 in 80 trials.
Assuming a 1/12 probability of a casualty, there is less than a 1% chance of getting 1 casualty or fewer in 80 trials. But assuming a 1/36 probability, there's a 34% chance of 1 casualty or fewer. So in that case, it looks to me as if it's operating correctly.
Of course, that's not to say a bug might creep in under other circumstances.
The fact that you're getting casualties in sieges is a bug, certainly. There can be lots of sieges in a game, and if it's possible that 2 of your 3 casualties came that way it may indicate that that's the only bug. That is, it may be applying the right probability, but over too large a denominator.