Measurement Before Control: A Broadly Applicable Principle for Better AI Policy