For 100 targets, 90 are 'correct', plus 20x civs per-target is 90/2100 or 4% real accuracy.
Say you use a model that's only 50% accurate and limit yourself to 10 civs per-target, you're at 50/1100 or 4.5% accuracy!
I guess my point is that no self-respecting datascient would release a 50% accurate model, let alone one used to make life or death decisions and yet, in the application of this model, decisions made by humans about its use has made it no better than doing exactly that.