A prediction does not act on a single compound in isolation. It sorts a set of candidates so that a team with limited money and time can decide which ones to carry into the next, more expensive stage and which to set aside.
Suppose a team holds eight candidates and can afford to advance only three. Each candidate arrives with a risk score from the model. If the team sets the flagging threshold low, only the very safest candidates survive and the three slots go to the lowest-risk compounds; if it sets the threshold high, more candidates remain eligible and the team must choose among them using other evidence.
What the learner should notice is that the threshold is not a property of the model. It is a property of the portfolio. A team with many candidates and a cheap next stage can afford to be strict, because discarding a good compound costs little when replacements are plentiful. A team with few candidates and an expensive next stage, or one that urgently needs a success, is better served by a permissive threshold, because a false negative — a doomed candidate that slips through — is the error it can least afford. The same model, the same scores, and the same compounds can justify different decisions in two different organizations.
A second thing to notice is that the prediction reorders attention rather than removing judgment. The candidates the model ranks lowest are not deleted; they are deprioritized, and a team that later gains budget or new evidence can revisit them.