Zheng Zhang

AI answers sound confident even when they are wrong. This verifier checks a claim the way a fact-checker would: it splits the claim into atomic facts with Muse Spark, fetches evidence triples from Wikidata for each fact, and judges each fact only from those triples. If the evidence is too weak, it says NOT ENOUGH INFO instead of guessing. Facts judged NOT ENOUGH INFO go to Muse Wonder for a second review. On a 100-claim FEVER sample, it scored 39/100: near-perfect on NOT ENOUGH INFO (30/33) with low recall on SUPPORTS/REFUTES, because retrieval — not the judge model — is the bottleneck. When the right triple is retrieved, the mechanism works: "Stadium Arcadium won Grammys" verifies via "award received | Grammy Award for Best Rock Album." All code was written during the hackathon; the repository is public with a runnable pipeline, tests, and honest scoring.