Fair that I didn't instruct it explicitly to make more pro-user choices, it just seemed to think slurping as much information into the backend was an default intention. Wasted a few more tokens to iterate on it to remove things, but it was IMO interesting enough that I finally submitted feedback around what I imagine is an interesting training problem.
