Really skeptical about the false positives/negatives. The ML they are using around "rich snippets" is pretty lousy. I can find lots of bad autogenerated answers in just a few minutes.

Or would this use some approach radically different, and more accurate[1], from what's driving rich snippets?

[1] well, as much as you can be accurate for subjective measures