There is a fairly well known discussion point in quant trading circles that says you should always bet against the end of the world. If the logic is not immediately obvious to you, it goes something like this: if the world doesn’t end, I make money, and if the world does end, it won’t matter that I lost money.
I don’t know what the conditional probability of the world being “destroyed” is (defined here loosely as lots more nuclear weapons detonate, massive environmental damage, all world governments collapse, money is worthless) if one major city is destroyed by a nuclear weapon, but I would think its at least 25%.
All this to say, if you are making any decisions based on the implied odds of that market, its probably rational to consider them as higher, since we are really looking at the odds of [nuclear weapon goes off AND the world doesn’t end].
Supposing, maybe not realistically, that one could build a reputation for both torturing POWs during wartime and being a reasonable power to surrender to, the torturing POWs might even be productive to ending a war quickly.
My own intuition is that warcrimes are largely in the shape of “gentleman’s agreements” that let people kill each other and wage war without mostly undue suffering, and powers involved are mostly happy to keep them in place because if both sides start to violate them life just gets worse for everyone without the needle of who is winning the war moving much.
As applied to torturing POWs: if faction A starts torturing people en masse, they might gain a very small edge in the immediate term in the war from intimidation or better intelligence from those POWs. They lose this edge rapidly once faction B starts doing the same, and now everyone is worse off from that new norm. Most warcrimes I can think of off the top of my head fit this mold, eg, killing civilians. It applies perhaps less so to things like destroying civilian infrastructure or conscripting children, in that you can construct lots of cases where one side is much more willing and able to do that than the other.