DOJ Epstein Files (File Transparency Act)

Scoring uses an investigative triage prompt. Inputs can be OCR text or rendered page images, and outputs follow a strict JSON schema consumed by this viewer.

• Evidence discipline: The model is instructed to use only explicit evidence in the file and avoid invented names, dates, agencies, locations, or relationships.
• Low-signal handling: Unreadable, decorative, duplicate, or otherwise weak-content pages are assigned low scores.
• Lead focus: Scores prioritize investigative value for concrete links between influential actors and misconduct, coercion, trafficking, financial flows, legal exposure, corruption, intelligence activity, or cover-up behavior.
• Entity normalization: Power mentions are canonicalized (aliases collapsed), and agency_involvement plus lead_types are constrained to controlled vocabularies.
• Output contract: Each record must return headline, importance_score, reason, key_insights, tags, power_mentions, agency_involvement, and lead_types.

The rubric is integer-based from 0 to 100. Practical interpretation: 0-20 low signal, 21-50 limited to moderate leads, 51-70 strong leads, 71-85 high-impact leads, and 86-100 exceptional to blockbuster leads.