Split fit from engagement and score them on separate axes. Fit is firmographic and mostly static; engagement is behavioural and decays. A single blended number hides which of the two drove the score, which is exactly the question a rep asks before working the lead.
Publish the score's components on the record so it is inspectable, and agree a re-scoring trigger. Negative scoring is worth adding only once the positive side is stable — otherwise you are tuning two moving parts at once. Set a conversion-rate target per score band so the model can be falsified.
Every claim above comes from work LeanScale published. These are the sources.