P. Barnett and L. Thiergart, “What AI evaluations for preventing catastrophic risk can and cannot do”, SI, vol. 2, no. 4, Jan. 2026.