New research quantifies what happens to retrieval accuracy and token cost when you approval-gate and version context instead of running raw similarity search.