SNAPSCRIBE
Built at Built with Opus 4.7: a Claude Code hackathon · Apr 21, 2026 · Remote

Lawyers and researchers lose careers when AI-fabricated citations slip into their published work. On April 18, 2026, Sullivan & Cromwell apologized to a federal judge for filing five fake AI-generated case citations. GPTZero scanned 4,841 NeurIPS 2025 papers and found 100 fabricated citations across 53 of them, surviving review by 3 to 5 expert researchers each. Truth Seeker catches every fake citation in any document before it ships. Stage 0 is a deterministic check across eight registries (CourtListener, CrossRef, DataCite, OpenAlex, Semantic Scholar, DBLP, arXiv, PubMed) — 800 ms, zero LLM calls. Stage 1 is a Claude Opus 4.7 agent on Anthropic's Managed Agents platform that reads the document with the native PDF skill, fetches every primary source into a single 1M-token window, and grades every claim against what's actually there. Stage 2 (Panel of the Dead) runs five Opus 4.7 personas in parallel for adversarial verification. Reproduced live on the actual Sullivan & Cromwell filing: 53 verdicts in 24 minutes for $7. All five admitted fabrications caught, 14 more they hadn't noticed, zero false positives on the corrected version. Apache 2.0, calibration-tested, MCP server included.