To me this comes across as word salad.
The edits are done based on reasoning, to accomplish a "target" improvement, that is itself reasoned and designed, taking into account a reasoned understanding of the existing functionality of the weights.
This is the furthest thing possible from some gradient optimization "over fitting" weights to tune a pedantic measure/target. That could be relegated to one-time greenfield-model bootstrapping.