The post introduces two measures intended to supplement dataset-level g-mean², which evaluates how well a monitor distinguishes target behavior from non-target behavior.
The pair-matched metric compares models only on shared task instances that support the relevant comparison. OpenAI says this reduces bias caused when different models generate valid positive behavior on different subsets of a task pool. For comparisons involving three or more models, the post proposes combining pairwise estimates with HodgeRank.
A second measure, post-control integration, simulates monitor-gated rejection sampling. It reports bad-output probability, overall deployment probability, and the fraction of deployed outputs that are bad, including the effect of retries after rejection. The authors also derive a pessimistic lower bound for intervention g-mean²; the previously used score is the corresponding optimistic upper bound.
The source does not report a product release or a change to OpenAI model access. It explicitly describes the deployment-control setup as hypothetical and says OpenAI does not deploy its current models in that way.