OpenAI researchers introduced two new metrics for AI monitorability. Pair-matched metric compares models on shared task instances to reduce bias. Post-control integration metric simulates monitor-gated rejection sampling. It estimates how often a system deploys bad outputs and how retries change this. They also derived a lower bound for intervention g-mean² to make causal attribution uncertainty explicit. The metrics address distribution mismatch and deployment risk gaps in g-mean². Pair matching isolates monitorability from capability-driven instance selection. HodgeRank ordering combines pairwise comparisons for multiple models. Synthetic experiments validate the metrics' ability to recover true monitorability orderings.