thresholds.json

{
  "probability_threshold": 0.9802263879740749,
  "margin_threshold": 3.903435230255127,
  "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"
}