Understand this story
On October 20, 2026, OpenAI published a post on its Alignment research blog proposing two complementary monitorability metrics in addition to g-mean2. According to the post, the pair-matched metric compares monitorability only on instances that are jointly supported by multiple models; the post-control risk metric simulates the probability that the system still produces harmful outputs after the monitor rejects an output. The post notes that g-mean2 was previously used to measure a monitor's ability to distinguish positive from negative samples, but because different models have different subsets of valid positive-sample tasks, the scores may reflect different mixtures of tasks. The post also states that if the monitor is used as a deployment control, the model can retry after being rejected, and the risk depends on the frequency of bad behavior, which outputs the monitor accepts, and the number of attempts allowed. This content currently comes from OpenAI's official blog and has not yet been cross-verified by other independent sources.
AI-generated from public reports
Latest development · Report
OpenAI proposes complementary monitorability metrics beyond g-mean²Developments
1 confirmed developments
The OpenAI blog proposes two complementary monitorability metrics
Source:English text is being prepared.
Reports and discussion timeline
Browse reports and discussion by time. Their evidence levels remain distinct.
Oct 20
OpenAI: Alignment Research BlogFirst-party
OpenAI proposes complementary monitorability metrics beyond g-mean²
OpenAI proposes two complementary metrics beyond g-mean²: the pair-matched metric compares monitorability only on instances jointly supported by multiple models, while the post-control risk metric simulates the probability that the system still produces harmful outputs after the monitor rejects outputs. g-mean² was previously used to measure a monitor's ability to distinguish positive from negative samples, but because different models have different subsets of valid positive tasks, scores may reflect different task mixtures. The article notes that if the monitor is used as a deployment control measure, the model can retry after being rejected, and the risk depends on the frequency of harmful behavior, which outputs the monitor accepts, and the number of allowed attempts.