A research lab is developing a language model to act as a medical diagnostic assistant. They are debating between two training approaches:

1. Outcome-based Supervision: Rewarding the model only when it provides the correct final diagnosis.
2. Process-based Supervision: Rewarding the model for each correct step in its diagnostic reasoning process (e.g., correctly identifying symptoms, listing potential conditions, and ruling out alternatives).

Evaluate the trade-offs between these two approaches, specifically in the context of this high-stakes medical application. Which approach would you recommend and why?