Brain of Zoron turns a client's source documents into briefs and master documents.
The risk with agents doing that work is that they quietly decide what's true.
Agents built on the Claude Agent SDK read the source documents one at a time and propose items for a canon in Postgres with pgvector.
Nothing enters the canon until a person approves it. Approval takes a token the agents never receive, and the MCP connector, deployed on Vercel, approves as the key's owner.
Briefs and master documents are rebuilt from the approved canon, never straight from the sources.
The answer-key document is refused by its hash, so the agents can't read the answers, and an eval harness grades them against a set of known facts.
Built in a week in October 2026, 91 commits, with UAT documents and an eval harness.
The approval lives outside the prompt, in a token the agents don't hold.
