Dev builds anti-hallucination research agent that verifies 108 citations word-for-word
ApprehensiveMark1308 · reddit · 2026-11-08
A Reddit developer shared a multi-agent research system (supervisor → retrieval → sandbox → synthesizer → critic) built to stop models from inventing facts:
- Word-for-word citation checks: every claim must point to a fetched-page snippet verified verbatim; unmatched claims fail.
- No model math: prices and metrics are computed in a sandbox, then recomputed by a critic.
- Facts vs. hypotheses kept separate: weak evidence yields "inconclusive" rather than a confident guess.
- 8 rule-based checks first, then an LLM judge; failures send the report back for revision.
A test on Stripe's public pages verified 108 citations, extracted 60 prices, and marked 2 of 3 hypotheses inconclusive — the outcome the author was happiest with. The project is open source (nexus-agent), and the author plans to spin the critic into a standalone library that any agent can call with an answer plus sources to flag unsupported claims.
More from coding & agent
- zml team hits "RSI": DeepSeek 4.1 Flash token use exceeds Codex+Claude combined — RemiCadene · 2026-11-08
- Agent S4 lets agents evolve reusable computer-use interfaces from their own task experience — burkov · 2026-11-08
- Running 50-100 Agents, Dev Weighs Local Decision Models to Cut Latency — f1yingbanana · 2026-11-08
- This PM replaced the PRD with an interactive Claude artifact — PawelHuryn · 2026-11-08
- Spacedrop lets you drag-and-drop code to deploy in seconds, no GitHub push needed — saheedniyi_02 · 2026-11-08
- Beam's technical report details its sandwich architecture — brandondamos · 2026-11-08