Snowie doesn't understand the concept of "1- alpha" or "minimum defence frequency", nor should it, because MDF is not GTO. (See multiple posts about this in the theory forum).
Snowie doesn't try to prevent its opponent from making money. It built its ranges by trying to maximise its
own EV, and it learned to do that by trial and error, until it had a list of hands that are or aren't profitable in situation X/Y/Z. Sometimes that means check-folding the river at
what looks like an "exploitable frequency". Sometimes it means hero-calling with king high.
Snowie doesn't care about 'theory', as such. It makes bets and calls because they were profitable when it did them while training. If you take "weird" lines with "weird" hands, it will often make what look like "weird" river decisions, possibly because it lacks the training for those weird situations. I mean, there are more possibilities in NLH holdem than there are atoms in the universe, so you can't seriously expect Snowie to know the optimal play for every single one. (Although the initial advertising for the software might have given this indication!)
Non-standard bet-sizes also confuse it. An early version of Snowie could be exploited by minbetting, because the "Snowie brain" treated that bet-size the same as it would a 1/2 pot bet, so it would fold to a minbet way more often than anyone with a human brain would.
There are definitely anomalies in its river ranges, which are most likely due to sample-size problems, but it's not like it happens every hand.
It used to frequently do weird/bad stuff on the
flop too, like folding a ten-high flopped flush (but calling with a 9-high flush) on some boards, perhaps because - during its initial training - every time T8s flopped a flush on a certain board, it ran into a better hand, which made the engine think that T8s should be folded in that spot. It's only from playing hundreds of millions more hands that it can correct the mistake.
If you spot some really bad mistakes/anomalies, the support team might be interested. I have no idea if/when the AI is going to be updated, but you can always email and ask.