[alm1]

that's why a lot of responders are calling the approach naive. This function is potentially the last layer in thousands of functions. Thinking that one can predictably change the behavior of a large nonlinear model via one variable is like thinking you can change the US political system by passing one good law in Reno, Nevada.

This approach only works if auditability is mandated to also be a property of the model, which typically reduces model complexity and accuracy too. We do this with credit scoring models semi-successfully.