Summary

Proposes minimum safety capability thresholds for AI model deployment based on empirical evidence from the corpus. Addresses what quantitative safety performance levels should be required before models are deployed in safety-critical contexts.