Anthropic privately pitched Congress a draft AI safety bill: blocks only for $1B or 50-casualty risks, via one court, with evaluators and preemption.