Nobody would say we shouldn't put a penal system into effect because it is wrong to harm innocent people for the greater good--even though harming innocent people for the greater good is exactly what it will do.

I am unsure how to map decisions under uncertainty to evidence about values as you do here.

A still-less-controversial illustration: I am shown two envelopes, and I have very high confidence that there's a $100 bill in exactly one of those envelopes. I am offered the chance to pay $10 for one of those envelopes, chosen at random; I estimate the EV of that chance at $50, so I buy it. I am then (before "my" envelope is chosen) offered the chance to pay another $10 for the other envelope, this chance to be revoked once the first envelope is selected. For similar reasons I buy that too.

I am now extremely confident that I've spent $10 for an empty envelope... and I endorse that choice even under reflection. But it seems ridiculous to conclude from this that I endorse spending $10 for an empty envelope. Something like that is true, yes, but whatever it is needs to be stated much more precisely to avoid being actively deceptive.

It seems to me that if I punish a hundred people who have been convicted of a crime, even though I'm confident that at least some of those people are innocent, I'm in a somewhat analogous situation to paying $10 for an empty envelope... and concluding that I endorse punishing innocent people seems equally ridiculous. Something like that is true, yes, but whatever it is needs to be stated much more precisely to avoid being actively deceptive.