I would like to see the word "intelligence" retired from this entire subfield until someone proposes a falsifiable behavioural test.

The review actually proposes three. Section 4. They are weak, but they exist.

The cost curve here is real but the pilot scale is 500 L. Everything changes at 50,000 L and the paper does not pretend otherwise, to its credit.

Anyone else notice that the reproducibility argument always ends with "we need better metadata" and never with "we need to fund replication"?

Because metadata is cheap to advocate for and replication is not.

Having done the 50k L version: foam control alone eats the margin they are projecting.