The model learns that hedging is a signal of lower-quality output. This creates a systematic bias toward sounding certain.
Nobody files a ticket that says “our architecture has an abstraction problem.” They file tickets saying the data is wrong, or ...