What it doesn't tell us though is how useful the rejected recommendations were.

Meaning, how many rejected solutions were sufficient to give the engineer enough context to turn a 30m task into a 5m task because they generated a recommendation, got an idea, rejected it, and rewrote it more efficiently or more correctly?

There's a lot of "devil in the details" likely buried in here.