Unlock: AI Control: Safety Without Assuming Alignment

AI control evaluates the whole protocol around a capable model against deliberate attempts to cause a prohibited outcome, instead of assuming the model shares the intended goals. This page draws the trusted boundary, compares protocol components, derives the violation rates of releasing versus deferring flagged work once the audit budget runs out, derives honest-mode task success under deferral, bounds risk across many dependent steps, and lays out what a control evaluation should report.