Simulated Control Evals Predict AI Safety With Lower Error
Control evals have so far been limited to purpose-built settings whose transfer to real deployment is questionable; SCEs let safety teams test safeguards against deployment-specific environments cheaply, without executing every task. The 3.6 vs 10.1 percentage-point RMSE gap against BashArena is the concrete benchmark to cite when arguing simulation fidelity. AI safety researchers and red-team builders can apply the technique by generating simulated tool responses for their own target environments.