{ "probability_threshold": 0.661988004118821, "margin_threshold": 0.6721660494804383, "recall_first_probability_threshold": 0.5, "rule": "flag as prompt injection if probability >= probability_threshold (equivalently margin >= margin_threshold)", "input_protocol": "flatten whitespace, tokenize with tokenizer.json (keeps its normalizer), keep the first 512 tokens" }