But this creates a strange situation where a paper presents detailed methodology and performance claims but code generating those results is withheld. The community can't verify or build upon the actual implementation.
The middle ground of sharing methodology without tools to reproduce seems to serve neither research nor commercial interests well.
If protecting IP is the priority, what's the incentive to publish at all?