The library VL database has PRU "% Return On Shareholders Equity" = 35% and net cash of $7.5Bn. Can't trust the data.

Their database certainly has errors, no doubt about it, but they aren't material very often.
I find that more often than not mismatches are a question of which statements were used, and what adjustments they make. (smoothing, skipping goodwill, whatever).
If (a big if) the mismatches are due to methodology rather than flat-out bugs, and the methodology is pretty consistent through the years, the historical results might be expected to keep some useful predictive power.

I would never put more than (say) 2% of a portfolio into a quant pick. And probably much less, these days.
As a rule of thumb, the allocation to a single carefully picked quant screen might be comparable in size to the allocation to a single carefully picked stock.

Jim