How an assessor could investigate alignment of unreleased AI models

Our proposal is to analyze a developer’s internal incidents and whether training may suppress evidence of worse misalignment.