compare-models-blindly

MEDIUM W011: Third-party content exposure detected (indirect prompt injection risk).

• Third-party content exposure detected (medium risk: 0.30). In scripts/blind_candidates.py, the runtime reads outsider-supplied JSONL from the provided input path (args.input), extracts two candidate fields and shared context, and writes the resulting free text into blinded-cases.jsonl for later LLM judging without any pre-selection by a specific item.