Deployment Simulation

Pre-deployment safety methodology: replay ~1.3M de-identified real conversations against a candidate model to predict its deployed behavior rates before release. OpenAI reports a median multiplicative error of 1.5x across 20 pre-registered behaviors, and that simulation reduces eval-awareness confounds from 99.72% (on standard evals) to 5.12% — a direct attack on the "models behave differently when they know they're being tested" problem.