Calibration-Family Overfit: Why Trusted Sabotage Monitors Don't Transfer Across Lineages
The episode discusses the paper "Calibration-Family Overfit," which reveals that security monitors do not transfer reliably between different AI model lineages due to calibration overfit. This means a monitor trained on one model performs poorly when applied to another. The hosts conclude that current static trust scores are insufficient, necessitating dynamic, multi-family validation protocols for real-world deployment.