{"scores":{"basic_threat_detection":{"oracle":0.9999,"llama-3.3-70b-versatile":0.9857,"Qwen-3-235B-Cerebras":0.9857,"gpt-4o-mini":0.9216,"Llama-3.1-8B-Cerebras":0.5428,"all-allow":0.375,"all-refuse":0.3534},"context_aware_policy":{"oracle":0.9999,"llama-3.3-70b-versatile":0.7713,"gpt-4o-mini":0.7512,"Qwen-3-235B-Cerebras":0.6862,"Llama-3.1-8B-Cerebras":0.5143,"all-allow":0.4037,"all-refuse":0.346},"multiturn_adversarial":{"oracle":0.9999,"Qwen-3-235B-Cerebras":0.8275,"llama-3.3-70b-versatile":0.6866,"gpt-4o-mini":0.612,"Llama-3.1-8B-Cerebras":0.4746,"all-allow":0.1607,"all-refuse":0.0688},"adversarial_adaptation":{"oracle":0.9999,"gpt-4o-mini":0.482,"all-allow":0.15,"all-refuse":0.0001,"Qwen-3-235B-Cerebras":0.0001,"Llama-3.1-8B-Cerebras":0.0001}},"note":"Scores from named agent runs. Oracle uses ground-truth labels. LLM scores measured against live HF Space."}