Comment by ByteAtATime

Interesting - I wonder if it's because coding doesn't use the specific token probabilities, while decision models do