Evaluation

Adversarial testing: short and long-context English

Red teaming of individual capabilities in high-risk categories using well-known published and unpublished techniques across single- and multi-turn conversations, including adversarial multi-turn automation similar to PAIR; attack techniques include multi-turn refusal suppression, hypothetical scenarios, personas/role play, disclaimers and warnings as response priming, gradual escalation