“Anthropic's safety measures for its models were inconsistent, applying classifier-based restrictions for some use cases and silent performance degradation for others.”