Is there a different way to identify the correct model from the Compose model?

Question

We are using 2.1 GA of Form Recognizer.
For every document variation, we create a separate project/model.
However, even with uniform training data set for each project and noticeable differences, we are still encountering issues where Compose model is not able to pick the correct model.
Based on observation, the main basis for picking the model is based on the Doc type Confidence.
However, it does not work as intended since it's picking the incorrect model.
May I know your thoughts on this issue?
Thanks.