Accountability
A dropped candidate fails invisibly. Because the outcome is never observed, the decision cannot be checked later, and the team cannot learn whether the model was right to flag it.
Feedback loops
Training data records past human choices, not a neutral sample of chemistry. If predictions push teams toward familiar compound types, future training data narrows further and the model's blind spots are reinforced.
Silent extrapolation
A model does not signal when a candidate lies outside its training territory. A score built on a thin comparison group is presented identically to one built on a rich group, so the apparent confidence can exceed the actual evidence.
Use the prediction as one input among several, keep a named person accountable for the decision, and stay alert to the cases where the model is being asked to judge something unlike anything it has seen.