precise-reducing-bias-evaluations

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

• Third-party content exposure detected (high risk: 0.75). The workflow ingests LLM-generated relevance judgments and confidence levels for top‑K documents (Step 1), which are free-form outputs produced by an external LLM judge rather than authored by the operating user, and these judgments are then read into the agent’s context for subsequent calibration/estimation.