Concord

Concord is an agent team that builds and enforces a glossary for long-form translation, the "shared brain" for a Chinese novel translated into English. Many translation tools now build glossaries automatically, but they don't show how each term was decided or measure whether the translation followed it. In Concord, Nemotron agents extract terms, group aliases (a nickname attached to its character), and decide each rendering through blind proposals, critique and a terminologist's written rationale, escalating hard cases to a human. Every decision is versioned. DeepSeek translates each chapter against the frozen glossary, GLM reviews, and code checks every chapter. Everything runs on Temporal: kill the worker mid-run and it resumes. The glossary grows chapter by chapter: each new chapter reuses frozen terms and debates only new ones, and when the agents gave the taxi driver the mentor's title, code caught the collision and a human decided. Measured on 3 chapters with the same writer model: the full system followed its glossary in every checked case (100% adherence, 0 banned variants, vs 82% for a one-prompt baseline), while the baseline gave the mentor three different titles. The whole run cost $0.125. All code was written today, with Claude Code as a coding agent; the README has the details.