When deciding my values, I want to be using an algorithm that would've gotten the right answer on slavery, even given 19th century inputs.

Even assuming that "19th century inputs" contained enough misinformation that this distinction makes sense, are you saying this independence of inputs should be a general principle? This actually seems sorta bad. A set of ethics that would give the same answer if I was raised in 1500 as it would if I was raised in 2000 also seems likely to give the same answer if I'm raised in 2500.

If you have to choose between pushing a button that will save a life with probability 0.99, and cost a life with probability 0.01, surely it's not unethical after the fact if you got unlucky.

Nope, still unethical. I mean, you can define ethics as a property of actions rather than as something people do if you want. But surely, at least, that "surely" has to go.