Indie

Leo Pimentel, ML researcher. If you run LLM-as-judge panels you’ve hit this: three models disagree, and you have to turn that into one decision. Averaging the scores hides the disagreement — you can’t tell “they agree it’s a 7” from “two said 9, one said 3.” Zydah keeps the disagreement. Each judge returns a belief mass over {pass, partial, fail} plus explicit mass on “don’t know.” Conflict between judges isn’t averaged out, it accumulates as ignorance, and you get an interval instead of a point. Wide interval → escalate to a human. Narrow and high → commit. It’s Dempster-Shafer theory with Yager’s combination rule, which is the part that makes conflict visible instead of normalizing it away.