How do you predict video performance before it's published? (our approach + where it breaks)

Building the "predict performance pre-publish" feature for CreatorGrowAi has been the hardest part so far — YouTube's algorithm is a black box, so we're combining historical channel performance, title/topic similarity, and comment sentiment trends instead of trying to reverse-engineer the algorithm directly.

It's directionally useful, not perfect. Curious if anyone here has tackled a similar prediction problem (recommendation systems, content scoring, etc.) — what worked, what didn't?