Make agent skills actually followable: score risk in one command, bind behavior with a contract, and prove every run with an auditable alignment record.