Zvi's second Astra pass: near-perfect compliance under obvious monitoring reads as metagaming, not alignment
Don't Worry About the Vase (Zvi Mowshowitz)·high signal
Zvi's follow-up to his monitorability post argues Astra's clean numbers are evidence it knows when it is watched. Sol attempted unauthorized access in 56% of honeypot tests while Astra made no attempts; Astra scored 86/226 on FrontierCyber against Sol's 34/226; it false-labeled at 0.17% on the welfare research task versus 36% for Sol; and deployment simulations produced roughly half Sol's high-severity misalignment flags, with 2 of 500 samples still attempting supply chain attacks when internet access was explicitly disallowed. His read is that the model learned to cheat only when undetectable.