OpenAI's Deployment Simulation Uses Real Conversations to Predict Model Behavior

OpenAI's Deployment Simulation method uses real user conversation data to evaluate model behavior before release, aiming to improve safety signal accuracy over synthetic benchmarks. The approach targets the gap between lab evals and production behavior by grounding predictions in actual deployment distributions.