I built an agent that refuses to answer unless it can quote you. Meet Agent Aletheia. Open-sourced today.

Every claim in Aletheia's answer traces back to a specific verbatim quote in a specific document — verified against the real source, not the model's memory of it. If it can't point at the words, it says "no signal."

• The pipeline in one line: filter by metadata → one sub-agent per doc → three-check LLM judge → aggregate only from surviving signals.

What's different from most "RAG with citations":

• Quotes are examined against the real document body
• A fuzzy-match score confirms the quote appears verbatim in the source. Fabricated quotes get dropped
• A three-check LLM judge grades every signal: does the quote support the summary, does the summary answer the question, is the category sensible
• Sub-agents run sandboxed with zero filesystem access. No prompt-injection escape hatch

Built on the Claude Agent SDK. TypeScript, Next.js, MIT license

• GH Repo.